Recent investigations have induced relevant advancements of imaging interferometry, which is becoming a viable option for Earth remote sensing. Various research programs have chosen the Sagnac configuration for new imaging interferometers. Due to the growing diffusion of this technique, we have developed a self-contained theory for describing the signal produced by triangular FTSs and its optimal processing. We investigate the relevant disadvantages of multiplexing, and compare dispersive with FTS instruments. The paper addresses some methods for correcting the phase error, and the non-unitary transformation performed by a Sagnac interferometer. The effect of noise on spectral estimations is discussed.
© 2010 OSA
Recent investigations have induced relevant advancements of imaging interferometry, which is becoming a viable option for Earth remote sensing [1,2]. Various research programs have focused the opportunity to adopt the common path configuration for new imaging interferometers under analysis. More specifically, the Sagnac triangular configuration has been adopted for the development of some sensor prototypes, as reported by [3–6].
The Fourier Transform Hyperspectral Imager (FTHSI) was the first spaceborne imaging interferometer that operated on board of the AirForce Research Laboratory (AFRL) – U. S. Department of Defence (DoD) technological satellite MightySat II.1 (Sindri P99-1) after its launch in 2000 [4,7]. Precursors of the FTHSI were the HyperCam and the IrCam developed by the Kestrel Corporation (USA) for airborne applications, the Spatially Modulated Fourier Transform Spectrometer (SMIFTS) developed by Hawaii University , the High Ètendue Imaging Fourier Transform Spectrometer (HEIFTS) by Science Application International Corporation (USA) , and the imaging interferometer developed by Applied Spectral Imaging (Israel) for laboratory applications . An additional scientific initiative that will exploit the use of a hyperspectral imaging interferometer (the ALISEO sensor  and ) is the MIOSAT mission of the Italian Space Agency (ASI). MIOSAT is a technological compact satellite (mass around 120 kg) that will host three payloads: a panchromatic camera, a MEMS Mach-Zehnder interferometer for atmospheric sounding, and the ALISEO sensor. Other remote sensing instruments adopting an interferometric approach have been reported in [11–14].
Some advantages are implicit in using imaging interferometers, due to their high signal (Jacquinot’s effect), and the option to adjust the sampled spectral range and resolution by changing the sensor sampling step and the instrument Field-Of-View (FOV) [2–6,15,16]. Critical points are instead connected with the high data-rate requested, the need to pre-filter the incoming radiation in order to avoid aliasing in the retrieved spectra, and the heavy data pre-processing for compensating instrument response and possible acquisition artefacts [17–20]. In this connection achieving new insights on the performance of imaging interferometers is an important topic that should originate relevant contributions to remote sensing applications and spectroscopic studies.
In this work we propose a theoretical analysis of Fourier Transform Spectrometry (FTS) with specific relations to the common-path Sagnac configuration. We gives a brief description of the optical configuration (Sagnac, without input slit) of the new ALISEO sensor, and discusses the theory describing its signal and the spectral estimations derived by it. The remainder of this paper is organized as follow. In Sect. 2 the main optical characteristics of the developed instrument are outlined. Sect. 3 discusses the theory of Sagnac common-path interferometers and describes a self-contained model of interferogram and spectral estimation. In this section we demonstrate that multiplexing usually produces a severe radiometric disadvantage in FTS (as compared with dispersive spectrometry), while the absence of exit slit (here termed as Fellgett’s advantage) is confirmed. We indicate some theoretical properties of the interferometric signal that can be useful for correcting the phase error originated by uncontrolled scanning offset. We also point out how this phase correction scheme interferes with experimental noise affecting the interferogram. We also find an analytical model describing the error introduced by the not unitary cosine-like transformation operated by the instrument, obtaining implicitly an algorithm for correcting its effects. Let us note that most of these findings are proved under very general assumptions, such that they represents as well the behaviour of almost each interferometer. Sect. 4 discusses the main consequences of our theoretical findings, while Sect. 5 summarizes the outcomes of our investigation, and sketches future work and open problems.
2. The Sagnac interferometer
Sagnac is an optical configuration for a class of interferometers that have a triangular ray-path, common for the two rays of the interferometer. Due to this reason this type of instruments are also called common-path interferometers, as opposed to tilted interferometers like the Michelson-Morley and Mach Zehnder instruments. Figure 1 shows the instrument’s optical layout, so describing the main characteristics of the Sagnac configuration. The light is first collimated by the lens L1, and travels the interferometer by means of a beam-splitter BS and two folding mirrors M1 and M2. Light emerging on the output port is then focused onto the output focal plane that holds an image detector by the lens L2. The instrument produces in its focal plane a stationary pattern of interference fringes of equal thickness that are localized at infinity. Let us note that interference fringes of a Sagnac instrument are sometimes termed as “equal inclination” fringes, as in Katzberg and Statham (1996) . However, equal inclination fringes are usually produced in Michelson-Morley interferometers, and are characterized by a non-uniform fringe distribution in the instrument focal plane, as discussed in . This feature is quite different from the behaviour of a narrow Field-Of-View (FOV) Sagnac instrument, in the focal plane of which fringe thickness is truly constant.
It can be shown that the BS provides part of phase-delay between the two interfering rays, the remaining part being originated outside the BS, and that the overall linearly changes with varying the angle of the entering ray with respect to the instrumental optical axis [9,10]. Due to the absence of entrance slit, the device acquires the image of an object superimposed to a fixed pattern of across-track interference fringes. Then, introducing a relative motion between the sensor and the object, each scene pixel exploits the entire interference pattern, hence its interferogram and spectrum can be assessed. We call this kind of optical layout “Leap frog” configuration. In remote sensing applications, each ground point is observed under several viewing angles while the sensor moves with respect to the target, so a 3-dim array of data (image stack) of varying phase offset is collected. This data-cube is first processed in order to extract the complete interferogram of every image pixel, then it is inverse cosine transformed to yield a wavenumber hyperspectral data-cube.
One of the main drawbacks of the Sagnac configuration is that the image at the output port is affected by relevant vignetting frequently. This phenomenon evidently influences the fringe visibility, and demands for complex data pre-processing necessary for restoring the correct radiometric level of the signal in the far wings of the interferogram.
3. Theory of a Sagnac imaging interferometer
In this Section an attempt is made to develop a comprehensive theory of an imaging interferometer operating in the common-path, triangular configuration. Nonetheless, the properties discussed in Sections from 3.1 until 3.6 have been deduced on such a quite general ground that they can be applied to almost any kind of Fourier Transform Spectrometer (FTS).
In a Fourier Transform Imaging Spectrometer (FTIS) the acquired physical information is the interferogram , that is the power of the interference pattern generated by the two rays at the position x in the focal plane of the instrument, as stated in the following relationship.Eq. (1), the wavenumber κ corresponds to the wavelength , and the integration over κ takes into account for a non-monochromatic light spectrum. Let us note that the depends on the inclination ϑ of the entering ray, and that the corresponding focal plane position x is the projection of ϑ by the equivalent focal length of the focusing element of the instrument.
We note that both rays pass throughout the BS two times, being subject to two attenuations due to internal extinction of the BS. For this reason the electric field amplitude attenuation factors and can be expressed using the power reflectance and power internal transmittance of the BS:9,10,27,32]:Eq. (4) we suppose that the position of the x origin exactly corresponds to the interferogram centre . The effect of an unknown offset of the interferogram centre will be addressed in Sect. 3.2. As can be seen from Eq. (3), the interferogram is an oscillating function, which reaches a minimum at at each wavelength in the typical Sagnac configuration. The interferogram also decays for large s where useful information regarding the target spectrum is contained.
Actually, the complete raw interferogram of the energy coming from a certain pixel of the observed scene is convolved by the pixel dimension, and sampled with a (square) grid whose extension is limited by the detector size D. This is true for a system equipped with a (2 dim) array detector, like a CCD or a CMOS sensor. Therefore, the sampled interferogram would be expressed as::22], and is proportional to the shah function adopted by Bracewell , and Walmsley et al . The ratio is the areal fill factor, a characteristic specific of the adopted detector. We define as the variation along the pixel area, and as the smallest variation between adjacent pixels. With this naming convention, the maximum in the sampled interferogram ( ) is given by:Eqs. (5) and (6) we suppose that the interferogram central sample always is measured, regardless of the actual b value.
Let us note that Eqs. (1) and (3) contain an integral transform that differs from the usual cosine transform, the difference being due to circumstance that the also depends on the wavenumber κ. Therefore and κ are not conjugate (canonical) variables and the integral transform performed by the instrument is not unitary. This point will be deeply addressed in Sect. 3.1. In order to adopt a coherent notation we indicate here as follows the integral transform performed by the device with the symbol , and reserve the symbol for the standard cosine transform. The following equation shows the basic relationship linking the to the operator:Eq. (5) would produce a spectral estimation modulated by the cosine factor , due to the translation . From a mathematical standpoint this modulation can be related to the fact that interferogram samples at lower s enter two times the inverse interferometer transform and their contribution is overweighted. On the other hand, it is possible to eliminate this cosine modulation simply applying the operator to the symmetrised sampled interferogram . Symmetrization can be made duplicating samples at large s on the opposite side of the interferogram. Such operation clearly requires that at least a short sequence of samples is collected on both sides of the interferogram, in order to roughly know the position of the interferogram centre (with an accuracy better than one pixel ). In summary we have:Eq. (9) also yields an important limit for the sampling frequency in order to avoid the aliasing phenomenon in the retrieved spectrum. As stated by Shannon’s theorem for ideal sampling [9,25], the chosen sampling frequency should be greater than the bandwidth of the concerned source, otherwise aliases corresponding to adjacent m values overlap each other. In our case it should occur that:24]), and Eqs from (5) to (10) are useful for introducing a clear nomenclature, and declaring the meaning of any symbols. Using Eq. (10), the minimum wavelength for which the original spectrum can be estimated without aliasing errors is . In other words, the shortest observed wavelength (greatest observed wavenumber) gives rise to a monochromatic fringe pattern that has to be sampled not less than two times per cycle along the entire interferogram. All wavelengths longer than will have their fringe cycles sampled by two detector elements at least and can be reconstructed. Supposing we haven’t aliasing effects ( ), it is possible to isolate the central alias ( ) by low-pass filtering Eq. (9) using an ideal filter . This yields:Eq. (11), a monochromatic source , i.e. a Dirac’s delta pulse, is estimated as a finite-width function, so worsening the available spectral resolution. The best spectral resolution ( ) can be written as:Eqs. (6), (10), and (12) it can be noticed that the parameters and are related each other by means of the asymmetry factor b, so that the interferometer overall performance mainly depends on instrument parameters like , M, and b. It is interesting to evaluate the interferometer performance while varying these parameters. Equations (12) can be rewritten as:Fig. 2 this dependency is plotted for an interferogram formed of 1024 samples (M) for two values of b parameter (0.5 and 1), for . The curve obtained for (quite asymmetric) shows the best interferometer performance but is unrealistic because symmetrisation requires a rough estimate of the interferogram centre, which can be assessed only relying on a sampling which overlaps both sides of the interferogram. Some samples have to be collected also on the “other” side of the interferogram making the parameter b less than unit. To this purpose it can be sufficient gathering two or three samples around the interferogram centre, so that b can approach the unit without reaching it. In this sense, the blue curve in the plot is anyway a realistic representation of the best interferometer performance.
From a theoretical standpoint the is related to the Geometrical Path Difference through the material refraction index . In the Sagnac configuration the has contributions from the raypaths inside and outside the beam splitter, which is made up of some glass or crystal having a spectrally variable refractive index. In our specific configuration, the path outside the beam splitter is in vacuum or air (for the airborne and laboratory prototypes), the refractive index of which is assumed to be unitary at all wavelengths. In view of Eq. (4) and considering these characteristics we write the general equations below:26]. The depends not only on the position x (i.e. the input angle ϑ), it is a function of the wavenumber κ too, hence κ and cannot be considered a couple of conjugate variables. Due to this property, the integral transform performed by a Sagnac interferometer is not a true cosine transform and it might not be invertible, meaning that the composite operator could not be unitary. In such a case the source spectrum estimate in Eq. (8) would be degraded by some unknown error that must be carefully evaluated.
3.1 Spectral dependence of the optical path difference
Let us investigate the properties of the and how it departs from the ideal unitary operator:Eqs. (14) one can write:Eqs. (17) can be fruitfully expressed using a power series of blocked to the first order:Eqs. (18) represents the derivative of the refraction index with respect to the wavenumber, and the solution of the ξ integral demonstrates that the input – output relationship of a Sagnac interferometer is affected by a wavelength dependent distortion given by:Eqs. (8) with the above result gives us the opportunity to write a thorough mathematical link connecting the true spectrum of the observed source with the filtered inverse interferometer transform of the sampled interferogram:27]. Let us note that the availability of an analytical formulation for correcting the effects of a spectrally dispersed may be relevant to many interferometric applications .
3.2 Phase error due to sampling offset
Perfect interferogram sampling is an infrequent circumstance, and in real applications no sample is exactly collected on the interferogram center. We term this phenomenon as “sampling offset”, and introduce the new symbol that indicates an unknown x translation of the sampling grid. Obviously, the sampling offset has to be less than the sampling step , otherwise its integral part has to be included in the asymmetry parameter for a coherent analysis of sensor performance. In the past, a great effort has been devoted to mitigate or remove the phase error introduced by uncontrolled sampling offset [24,28–30]. As a consequence of offset, no sample is exactly matching the interferogram center, introducing a phase error in the source spectrum estimation computed as inverse complex interferometer transform . The phase term disturbance can be removed taking the norm of the inverse complex transform , an operation that restores the result in the right hand-side of Eq. (20). This type of correction was devised by Connes , and it was investigated in depth by Walmsley et al . In the past, the main shortcoming of this correction scheme was its requirement for a significant computation burden that prevented its application.
Some authors  have hypothesized that noise affecting the interferogram sampling may interfere with phase correcting procedures, however the effect of noise on the previous phase correction scheme has never been clarified. As a matter of fact, the interferogram measurement being processed is affected by electronic and quantization noise that without loss of generality may be assumed to be a stationary zero-mean Additive White Noise (AWN). The first problem is that the noise degrades the measured interferogram but we are interested to its effects on the spectral estimation . A properties of stochastic processes and random fields is that the Fourier Transform of a stationary white noise field (process) again is a stationary white noise field (process) with the same spectral power density (standard deviation) and null-mean . Let indicate the measurement noise affecting the interferogram, and its Fourier transform. According to , we can write:Eq. (20), as shown below:Equation (23) shows that the average of the inverse interferometer Fourier Transform is biased by noise, as reported in the second of Eqs. (23). This bias is due to the combination of the square modulus with the ensemble average operator that acts on the random component like an estimator of autocorrelation at zero-lag. As known, this zero-lag autocorrelation obtains the square modulus of the mean plus the variance. Therefore, the method of correcting the phase error by taking the modulus of the complex Fourier transform of the interferogram is well suited when the noise amplitude is negligible (low noise sensors), or whereas the noise standard deviation of the detector is calibrated with an independent measurement and removed from the final result. Let us note that only the noise components that are independent of the source (e.g. quantization error and electronic noise) can be reliably calibrated and removed. Moreover, photonic noise should directly affect the symbol in Eq. (23), so the spectral radiance in these relationships should be considered as an additional stochastic field having mean , and the standard deviation characteristic of photonic noise. Also for this more accurate modeling of the combination of the square modulus with the ensemble average operator behaves like an estimator of autocorrelation at zero-lag. Hence, it is easy showing that this improved modeling of photonic and electronic noise would lead to a final result quite similar to that shown in the second of this Eq. (23). Indicating the standard deviation of photonic noise with the symbol , it results:33] (e.g. non-white noise).
Considering that the bias affecting spectral estimations can partially removed with a simple calibration measurement, the corrected estimator is always preferable because it prevents possible phase errors. It can be affirmed that the analyzed phase correction scheme provides reliable spectral estimations independent of possible sampling offset errors at the price of enhancing the weight of random noise and a possible residual estimation bias originated by photonic noise contribution (source dependent) that can’t be calibrated and removed.
3.3 Effects of the direct term
As shown in Eq. (3) the interferogram contains a constant contribution independent of x and proportional to the half power of the observed source. This contribution does not bring information about the source spectrum, and should be considered as a disturbance that may degrade the interferogram and the associated spectral estimations. The above reasoning is summarized in Fig.s (3) and (4). Figure (3) indicates the informative part of the interferometric signal (the small amplitude ripple of the orange curve) and the useless part of the interferogram: the dc term plotted as a green straight line. In Fig. 4 we show the inverse Fourier Transform (spectra) of the interferogram components already shown in Fig. 3.
Often, this direct term is removed before the inverse transformation that leads to the source spectrum estimation, although the presence in the sampled interferogram of a residual direct term does not affect the spectrum estimation significantly in many spectroscopic applications. In order to shows this property let us consider the presence of an additional constant term in the interferogram. Due to this constant Eq. (23) changes as stated by the following equation:Eq. (24) follows straightforwardly. The perturbation is negligible for wavenumbers κ greater than , where the function on the right hand-side of Eq. (26) has its first null. The condition for which the perturbation is negligible defines an interval that evidently limits the operating spectral range of the interferometer, as stated by the following relationships:Eqs. (23) or (24). The sole exception is constituted by those cases in which the overall sensor response is not properly compensated, apodization is applied before inverse transformation, or uncontrolled vignetting introduces a broader bell-shaped profile whose convolution with the pulse in Eq. (26) gives rise to a disturbance that might interfere with spectral estimations (significantly lowering the limit).
3.4 Fellgett’s advantage of Fourier Transform Spectrometers
Fellgett’s advantage should be connected to the circumstance that each interferogram sample (pixel) benefits of radiative contributions from any wavelengths of the observed source [34–38]. In  Fellgett writes: “I recognize that a major inefficiency in spectrometry in the infrared region (...) is that the available observing time has to be shared among all the observed spectral elements. This inefficiency can be overcome by multiplexing all spectral elements through a single detector; that is to say, imposing mutually orthogonal modulations on the separate elements, and sorting out their individual contributions in the final output”. In such a way the overall physical signal level outputted by an interferometer should be much higher than that achieved by a dispersive instrument (i.e., grating). Fellgett idea was that the multiplexing permitted the observation of all the spectral elements at once even using a single detector, as in , where he writes: “Multiplexing is specifically associated with orthogonal sets of functions, but it is not at all necessary that these should be the trigonometric functions as in Fourier spectrometry.” This effect has received various interpretations in the past. As an instance the multiplexing advantage has been associated in  to the lack of output slit in interferometers, a rather different representation that may be also related to the property of Fourier Transform multiplexing of having poor and unimportant spectral dispersion in its output focal plane. Surely, the advent of 1-dim and 2-dim detectors (e.g. CCD or CMOS devices) permitted a similar advantage even for traditional dispersive spectrometers making Fellgett’s advantage outdated. Other authors has pointed out that summation of independent interferogram samples implicit in performing the inverse cosine transform should reduce the effect of incoherent noise on the spectral estimations (e.g .). Often, Fellgett’s advantage has been stated in term of an increased SNR available in interferometers, a phenomenon that should be more evident when the main noise source is uncorrelated and due to the detector .
Evidently, the multiplexing effect has been misinterpreted in the past frequently, also assuming that a higher spectral resolving power can be achieved by a Fourier Transform Spectrometer avoiding the effect typical of dispersive spectrometers where a higher spectral resolution implies the measurement of a fainter signal. It is easy to demonstrate that the above apparently accurate reasoning is very often untrue for a generic FTS that obeys Eq. (3). Hence, it is interesting stating on more firm basis (quantitatively) possible radiometric advantages of FTS, stemming from a physical and mathematical investigation of the signal produced by FTS instruments.
In the following we assume that Fellgett’s advantage is associated to the absence of exit slit, while possible radiometric advantages connected with the typical FT multiplexing are termed multiplexing advantage. This definition of possible radiometric advantages of FTS help us to include in the following analysis new features of modern detectors that weren’t available sixty years ago (e.g. array detectors). Let us note that in many works on interferometry the phrases “Fellgett advantage” and “multiplex advantage” are used synonymously. However, our purpose is not getting a better representation the original thinking of Fellgett or other authors. We arbitrarily assume the above definitions for sake of simplicity, aiming at obtaining a clear discussion of their implications.
3.4.1 Noise and multiplexing advantage
Here as follows we will make use of continuous representations of signals involved in the interferogram and spectral estimations, avoiding the mathematical complexity of handling with sampled signals. Any results are easily extended to the case of a sampled signals. Let us write the single-sided interferogram signal model as:12] write: “Fourier Transform Spectrometers, on the other hand, observe a linear sum of elements of all portions of the input spectrum. When reconstructed, the presumably uncorrelated noise samples combine incoherently, while the signal adds coherently”. This claim is erroneous because it applies a property of the arithmetic addition of uncorrelated random samples to the FT of the same random sequence. As shown in Eq. (22), and in many textbooks on random fields and stochastic processes (e.g .) the FT of a random sequence obeys a different law, and the signal-to-noise ratio (SNR) is an invariant quantity with respect to Fourier transformation. For the interferogram of Eq. (23), the average variance of an auto-uncorrelated noise can be estimated as:Eq. (29) represents a common form of noise amplitude estimator, adopted in many works. Interestingly, this equation holds true in strict sense if the involved noise term is at least a variance-ergodic field (process). The variance ergodicity simply requires an auto-uncorrelated normal field. Evidently, any form of white noise or of auto-uncorrelated normal noise is variance ergodic. The SNR in the interferogram domain is given by:Eqs. (31) and (32) show the average SNR of the interferometer physical signal in both domains, but that they do not represent the effective SNR of interferometric measurements. As shown in Section 3.3 the signal (and its ) has a large amplitude due to the presence of a non-informative component, which beautifies the total SNR.
3.4.2 The constant term in the interferogram
The measured interferogram takes its high amplitude from a useless contribution represented by the first term in the right hand-side of Eq. (28). This term is not useful because it does not bring information concerning the source, with exception of its panchromatic energy. The truly source-informative signal in the interferogram is a tiny undulation around the constant term . This behavior is evident when we consider the inverse FT of the constant term: it originates a pulse located that doesn’t add information about the source spectrum, a circumstance that has been shown in Eq. (26).
It is worth noting that Plancherel theorem implies that the integrated power carried by the informative component of the interferogram equates a quarter of the integrated power carried by the source spectrum, ideally measured by a dispersive instrument.
3.4.3 The informative part of the interferogram
In this Section we will show that multiplexing by trigonometric orthonormal functions originates a serious radiometric disadvantage connected to the circumstance that the power carried by the effective signal in the interferogram is concentrated at lower s. In other words, spectral information concerning the source at high resolution (the finest spectrum details) is held in the subtle tails of the interferogram . In order to elucidate this point with an example let us consider a radiation source having a rectangular spectrum with bandwidth B as indicated in the following equation:
It is evident that using a dispersive spectrometer at the generic spectral resolution we need a radiometric resolution finer (better) than the power transmitted by this frequency range: . When considering a Fourier Transform Spectrometer (having the same scanning mechanism) the same spectral resolution can be achieved if the maximum observed by the device is not less than . And this requires the observation of a tiny interferogram oscillation whose amplitude is promptly deduced by the sinc term in Eq. (34).
It is worth noting that the assumption of a rectangular source spectrum is rather unrealistic, and should be replaced by a generic function continuous and Lebesgue integrable with its first l derivatives. The requirement that the lth derivative is Lebesgue integrable and admits a finite and invertible Fourier Transform means that the interferogram ripple (i.e.: the Fourier transform of ) dies away for large enough s at least as rapidly as . In such a case the radiometric sensitivity requested to a FTS device for reaching the spectral resolution is:Equation (37) is straightforwardly deduced applying the Reimann-Lebesgue lemma to the Fourier Transform of the lth derivative, as shown in . Let us note that we have assumed the interferogram to be a function of the product , a choice that conveniently accounts for the uncertainty principle. The term is an unknown factor of the asymptotic limit of the transform of the source spectrum . The above asymptotic behavior clearly shows that the previous estimation of held in Eq. (36) is the most favorable condition for an interferometric measurement, while for a realistic case we obtain from Eq. (37) the following comparison:Eq. (34) are examples of functions non-integrable in the sense of Lebesgue that, however, admit convergent Fourier integrals (or a generalized Fourier transform). Typical examples of source spectra that are Lebesgue integrable are the Gaussian and the Lorentz profiles. In summing up, Eqs. (37), and (38) mean that the radiometric accuracy requested to an interferometer for achieving the spectral resolution is usually higher than the one required to a dispersive spectrometer with the same performance. The above results can be condensed in the following theorem.
Lemma 1. Given a FTS with a radiometric accuracy matched in the sense of Eq. (38) to the desired spectral resolution , and assigned a dispersive spectrometer having a radiometric accuracy sufficient for obtaining the same spectral resolution , then a (high) spectral resolution limit exists such that:Eq. (38)). The threshold can’t be foreseen on theoretical basis since its value depends on the spectrum of the actual source, i.e. its bandwidth.
3.5 Interferogram quantization and noise effects
In this Section we analyses the possible benefits of using interferometric and dispersive techniques. Since we have shown that the average SNR is the same in the interferogram domain and in its spectral estimation, performance of FTS and dispersive instruments can be computed in the domain in which the measurement is executed (interferogram for the FTS and spectral domain for dispersive devices).
The to ratio is an important estimate of the radiometric accuracy required to the interferometer, in order to reach the spectral resolution . From Eq. (38) the following relationship is easily obtained:Equation (40) makes even more evident the possible disadvantage of using an interferometer for high spectral resolution measurements when .
Stemming from Eq. (40) it is possible to derive a simple relationship that states the minimal quantization accuracy (number of bits ) requested to obtain the spectral resolution from interferometric observations:
Evidently, Eqs. (40) and (41) implicitly assume that the standard deviation of the overall noise affecting the interferogram measurement is less than the small amplitude ripple to be observed. In a different wording, the maximum digitized (the ) has to be less than the value where the envelope of the informative signal equates the noise level :
Interferogram samples collected at s greater than are meaningless and don’t add useful information but mainly noise to the spectral estimation. This elementary concept is illustrated in Fig. 5 where the blue curve represents the envelope of the informative component of the interferogram and the black dotted line indicates the noise amplitude . Remembering that the effective average SNR is the same in both domains for FTS, adding interferogram samples in which the signal amplitude is lower than that of noise results in abating the average SNR of the subsequent spectral estimations.
It is well known that noise affecting radiometric measurements comes from three main sources: quantization noise (the round-off error), photonic noise, and instrumental noise. Noise caused by quantization and electronic detector and circuitry is not strictly connected to the optical configuration of the spectrometer under examination, and we can suppose that they affect in the same way the SNR of dispersive and FTS measurements. Therefore, we will focus our attention mainly on the photonic noise when investigating the advantages or disadvantages of FTSs. Photonic noise instead originates effects that are specific of the interferometer optical configuration and further degrade the performance of FTS. The point here examined regards the informative component of the interferometer signal, the only component that contains source spectral information. While the constant term doesn’t add contribution to the estimation of the source spectrum, the photonic noise originated by it strongly affects any spectral estimations. And this photonic noise contribution can’t be separated from the informative component of the signal. Moreover, since the non-informative term holds most of the power of the measured signal, this photonic noise contribution might be overwhelming at high spectral resolution (large s) with respect to the tiny informative component of the signal.
The photon flux experimented by an interferometer is related to the interferogram intensity that, according to Fig. 3 and Eq. (3), is mainly determined by the constant term . The photon flux is known to obey Poisson’s statistics law, giving rise to a flux variability whose standard deviation is just the square root of the flux itself. The standard deviation of the photon flux impinging in a unitary time interval over a sensor having unitary equivalent surface area within the narrow spectral interval around the generic wavenumber κ can be written as:Equation (45) allows us to write the maximum effective interferometer signal-to-noise ratio allowed by the photonic noise only:Eq. (38) we can write the asymptotic behaviour of the maximum signal-to-noise :
Due to the above behavior, measurements performed in the visible spectral range where the photonic noise has a large impact can’t be executed at high spectral resolution with a FTS. Nevertheless, in the infrared spectral range the photonic noise is mitigated by the lower photon energy, and high spectral resolution interferometric measurements are a viable alternative to dispersive instruments.
3.6 Lost samples during data collection or transmission
With the phrase lost samples we indicate measurements that, due to a transmission or acquisition error, assume a trivial (not informative) value, e.g. null. When such a circumstance occurs in the measurement performed with a dispersive spectrometer the concerned spectral channel has definitively gone. In general this lacking channel can’t be in any way recovered by interpolation or other predictive modeling, unless a large error is tolerated. The important point is that the resulting error (information lack) is concentrated in a single spectral channel that becomes useless.
The corresponding situation with FTS devices is quite different and less disruptive. After the inverse transform procedure that leads to the spectral estimation, the effect of a missing interferogram sample (error) is spread as a small amplitude perturbation over all the interpolated spectral channels. Moreover, the corresponding spectral error is quite low in any spectral channels, and no channel is missing. This is an advantage typical of FTSs and all those spectrometric devices which impose mutually orthogonal modulations to the spectral radiance impinging on separate detector elements (multiplexing).
The following list recaps the aspects relevant to the power efficiency of an generic FTS as compared to the traditional dispersive techniques.
- 1. The multiplexing measurement approach doesn’t produce advantage in reducing the experimental noise, nor the final SNR is optimized by the inverse FT operation.
- 2. The interferometric physical signal has a power high above that of a grating instrument observing the same source in the same experimental conditions. Most of this power is carried by a non-informative component of the interferometric signal (a constant term);
- 3. Using the Plancherel theorem it can be shown that the power held in the informative part of the interferometer signal (the raw interferometer signal minus the constant factor) is on average the same as the power of the spectrally dispersed signal available for a dispersive spectrometer;
- 4. Unfortunately, the cosine-like transformation operated by the FTS concentrates this equal power level (in the informative component of the signal) at lower s, strongly reducing the power efficiency of the FTS with respect to dispersive spectrometers when high spectral resolution measurements are performed (large s).
- 5. The interferometer has to measure a much more tiny signal as the dispersive spectrometer does, but this signal is found to be superimposed on a high-amplitude continuous radiation plateau. From a practical point of view, this feature represents a shortcoming since most of the digitalization accuracy of the employed detector is used to sense this useless (non informative) high signal.
- 6. This high-amplitude constant signal level originates a photonic noise of large standard deviation, so that the effective average SNR available in interferometric measurements is usually significantly lower than that obtained in dispersive observations with high spectral resolution. Decreasing is caused by both: the reduced level of effective signal amplitude at large s (see lemma 1) and the noise increase due to the augmented photonic noise (large amplitude of the non-informative interferogram component).
The above outcomes demonstrate that multiplexing is a radiometric disadvantage real for Fourier Transform Spectrometers. This characteristic was not fully recognized in many works about interferometry due to a misinterpretation of the noise role and the effect of the constant term. Fellgett’s advantage is maintained in FTS as long as it is defined as the signal amplitude increase connected to the absence of exit slit. Today, this advantage no longer is a prerogative of FT spectrometers due to the availability of low-cost array detector. Similar conclusions has been drawn by Schumann and Lomheim . The following Table 1 resumes the main differences among the traditional dispersive spectrometers and the most recent triangular interferometers that have been examined in this paper.
Furthermore, we point out that:
- • Sources having a spectrum with almost rectangular profile can be observed by a FTS that would closely match the optimal radiometric- resolution / spectral resolution compromise achieved by dispersive instruments;
- • Sources observed in the Infrared spectral range may show a spectrum closely approximated by a negative exponential function of the wavenumber. It can be shown that the Fourier Transform of such kind of spectra approach a Lorentz function of the . In such a case the radiometric accuracy required to interpolate the source spectrum with spectral resolution follows a rather favorable law that can be reliably measured with a FTS;
- • In the TIR spectral range the available signal contains radiative contributions from the optical elements along the ray path requiring cooling of the whole instrument. Usually, an FT spectrometer has lesser optical losses than a dispersive instrument, and sometime a minor number of optical surfaces along the raypath. Hence, the source-informative signal may be higher in FT spectrometers and be degraded at a lesser extent with respect to a dispersive instrument. This difference may mitigate the needing for cooling in spaceborne devices, originating a possible advantage of FTS in the TIR.
- • In the Visible spectral range, where due to the photonic noise the FTS performance is weakened, Fourier Transform Spectrometers can be advantageously employed for conducting low and medium spectral resolution observations of a broad-band source. In fact our theoretical modeling of FTSs performance of Eq. (37) provides us with a prediction of the required radiometric accuracy only as an asymptotic approximation valid for large s. Hence, in spectral applications where the maximum observed is not huge, the FTS technique can be a viable tool.
- • When observing a narrow-band source (e.g. a laser source with ), it is possible to perform high spectral resolution measurements while maintaining a mild radiometric resolution of the FTS even for applications covering the visible spectral range.
Possible applications of multiplexing advantage potentially remain confined to instruments adopting a non trigonometric set of orthonormal functions for representing the observed spectrum. Our theoretical investigations only exclude that this advantage holds true for Fourier Transform Spectrometers.
In this paper a deep theoretical analysis has been performed about the principles of Fourier Spectrometry pertaining the Sagnac triangular configuration. A special effort has been devoted to study the cosine-like instrument transformation in order to take into account the circumstance that the Optical Path Difference function depends also on the wavenumber. This investigation led us to introduce the concept of group , which expressly contains the spectral dispersion law of the material composing the beam-splitter (i.e., the essential element which originates by amplitude splitting the two interfering beams). It has been shown that the phase to group ratio is the amplitude of the non-unitary transformation operated by the instrument. This theoretical result may be useful for correcting the spectral estimations obtained with a FTS or for other interferometric applications.
An analytical expression of the sampled interferogram has been found, which includes the effects of on-pixel integration, spatial sampling offset, truncation of the interferometric measurement, and additional instrumental effects such as fringe visibility. Attention has been devoted to the effects of a null-mean additive white noise affecting the interferometric measurement, and to signal degradation produced by the photonic noise. If the complex inverse (Fourier) transform estimator is adopted in lieu of the standard cosine-like operator (for an instance, in order to automatically remove the phase distortion in spectral estimations due to the sampling offset error), the retrieved spectrum is affected by a bias proportional to the noise standard deviation, and subject to an enhanced noise amplitude (40% relative increment). We have shown that this behavior is introduced by the fact that the basic cosine-like inverse transform selects only the even part of the noise record.
Finally, a new interpretation of the so called Fellgett’s advantage has been discussed based on an analytical comparison of system performance (estimation of amplitude of the effective signal and Signal-to-Noise ratio) between the interferometric technique and the traditional dispersive spectrometers. This discussion has led to the conclusion that Fellgett’s advantage has been misunderstood frequently in the past. We have proved that the informative tail of the interferogram to be resolved requires a radiometric resolution much finer (depending on the illumination source’s bandwidth) than that is needed for a dispersive spectrometer operating at the same high spectral resolution. This circumstance also reveals the advantage of dispersive techniques of an exponential factor of pertaining the maximum Signal-to-Noise ratio with respect to the interferometric technique. In the Visible spectral range, where due to the photonic noise the FTS performance is weakened, Fourier Transform Spectrometers can be advantageously employed for conducting low and medium spectral resolution observations of a broad-band source. On the other hand, when observing a Visible narrow-band or an Infrared broad-band source, it is possible to perform high spectral resolution measurements while maintaining a mild radiometric resolution of the interferometer.
This work was carried out with the support of the Italian Space Agency (ASI).
References and links
1. M. J. Persky, “A review of spaceborne infrared Fourier transform spectrometers for remote sensing,” Rev. Sci. Instrum. 66(10), 4763–4797 (1995). [CrossRef]
2. B. Harnisch, W. Posselt, K. Holota, H. O. Tittel, and M. Rost, “Compact Fourier-transform imaging spectrometer for small satellite missions,” Acta Astronaut. 52(9-12), 803–811 (2003). [CrossRef]
4. L. J. Otten, A. D. Meigs, B. A. Jones, P. Prinzing, and D. S. Fronterhouse, “Payload Qualification and Optical Performance Test Results for the MightySat II.1 Hyperspectral Imager,” Proc. SPIE 3498, 231–238 (1998). [CrossRef]
5. P. G. Lucey, T. Williams, K. Horton, C. Budney, J. B. Ratfer, and E. T. Risk, “SMIFTS: A cryogenically cooled spatially modulate, imaging, Fourier transform spectrometer for remote sensing applications,” Proceeding of the International Conference on Spectral Sensing Research, Vol. 1, 251 – 262, (1992).
6. Y. Ferrec, J. Taboury, H. Sauer, and P. Chavel, “Optimal geometry for Sagnac and Michelson interferometers used as spectral imagers,” Opt. Eng. 45(11), 115601-115606 (2006). [CrossRef]
7. L. J. Otten, R. G. Sellar, and J. B. Rafert, “MightySatII.1 Fourier transform hyperspectral imager payload performance,” Proc. SPIE 2583, 566–575 (1995). [CrossRef]
8. P. G. Lucey, K. A. Horton, and T. Williams, “Performance of a long-wave infrared hyperspectral imager using a Sagnac interferometer and an uncooled microbolometer array,” Appl. Opt. 47(28), F107–F113 (2008). [CrossRef] [PubMed]
9. A. Barducci, P. Marcoionni, I. Pippi, and M. Poggesi, “Simulation of the Performance of a Stationary Imaging Interferometer for High Resolution Monitoring of the Earth,” Proc. SPIE 4540, 112–121 (2001). [CrossRef]
10. A. Barducci, A. Casini, F. Castagnoli, P. Marcoionni, M. Morandi, and I. Pippi, “Performance assessment of a Stationary Interferometer for High-Resolution Remote Sensing,” Proc. SPIE 4725, 547–555 (2002). [CrossRef]
11. M. Bliss, “Demonstration of a static Fourier transform spectrometer,” Proc. SPIE 3541, 103–109 (1999). [CrossRef]
12. S. J. Katzberg, and R. B. Statham, “Performance Assessment of the Digital Array Scanned Interferometers (DASI) Concept,” NASA Technical Paper 3570, August 1996.
13. P. D. Hammer, F. P. J. Valero, and D. L. Peterson, “An imaging interferometer for terrestrial remote sensing,” Proc. SPIE 1937, 244–255 (1993). [CrossRef]
14. S. Subramaniam, B. Y. Ravindra, B. Rabindranath, B. G. Basheerullah, P. V. Viswanath, and O. P. Bajpai, “Stationary spatially modulated fourier transform spectro-radiometer,” J. Indian Soc. Remote Sens. 31(3), 187–196 (2003). [CrossRef]
15. R. F. Horton, “Optical Design for High Ètendue Imaging Fourier Transform Spectrometer,” Proc. SPIE 2819, 300–315 (1996). [CrossRef]
16. D. Cabib, R. A. Buckwald, Y. Garin, and D. G. Soenksen, “Spatially resolved Fourier transform spectroscopy (spectral imaging): a powerful tool for quantitative analytical microscopy”, in Optical diagnostics of living cells on biofluids,” Proc. SPIE 2678, 278–291 (1996). [CrossRef]
17. J. Genest, P. Tremblay, and A. Villemaire, “Throughput of tilted interferometers,” Appl. Opt. 37(21), 4819–4822 (1998). [CrossRef]
18. P. Jacquinot, “The luminosity of spectrometers with Prisms, Grating, or Fabry-Perot Etalons,” J. Opt. Soc. Am. 44(10), 761–765 (1954). [CrossRef]
19. M. R. Descour, “The Throughput Advantage In Imaging Fourier-Transform Spectrometers,” Proc. SPIE 2819, 285–290 (1997). [CrossRef]
21. F. A. Jankins and H. E. White, Fundamentals of Optics, (Mcgraw-Hill College; 4th edition, 1976).
22. W. Goodman, Introduction to Fourier Optics, (McGraw-Hill, New York, 1968).
23. R. Bracewell, The Fourier transform and its applications, (McGraw-Hill, New York, 1965).
24. D. A. Walmsley, T. A. Clark, and R. E. Jennings, “Correction of off-center sampled interferograms by a change of origin in the fourier transform; the important effect of overlapping aliases,” Appl. Opt. 11(5), 1148–1151 (1972). [CrossRef] [PubMed]
25. M.-L. Junttila, J. Kauppinen, and E. Ikonen, “Performance limits of stationary Fourier spectrometers,” J. Opt. Soc. Am. A 8(9), 1457–1462 (1991). [CrossRef]
26. R. L. Hilliard and G. G. Shepherd, “Wide-Angle Michelson Interferometer for Measuring Doppler Line Widths,” J. Opt. Soc. Am. 56(3), 362–369 (1966). [CrossRef]
27. P. Hlubina, D. Ciprian, J. Lunacek, and R. Chlebus, “Phase retrieval from the spectral interference signal used to measure thickness of SiO2 thin film on silicon wafer,” Appl. Phys. B 88(3), 397–403 (2007), doi:. [CrossRef]
28. L. Mertz, “Auxiliary computation for Fourier spectrometry,” Infrared Phys. 7(1), 17–23 (1967). [CrossRef]
29. H. Sakai, G. A. Vanasse, and M. L. Forman, “Spectral Recovery in Fourier Spectroscopy,” J. Opt. Soc. Am. 58(1), 84–90 (1968). [CrossRef]
30. M. L. Forman, W. H. Steel, and G. A. Vanasse, “Correction of Asymmetric Interferograms Obtained in Fourier Spectroscopy,” J. Opt. Soc. Am. 56(1), 59–63 (1966). [CrossRef]
31. J. Connes, “Recherches sur la spectroscopie par transformation de Fourier,” Revue d’Optique 40, 45–265 (1961).
33. A. Papoulis, Probability Random Variables, and Stochastic Processes, (McGraw-Hill International Editions, Third Edition, 1991).
34. P. R. Griffiths, H. J. Sloane, and R. W. Hannah, “Interferometers vs monochromators: separating the optical and digital advantages,” Appl. Spectrosc. 31(6), 485–495 (1977). [CrossRef]
35. R. J. Bell, Introductory Fourier Transform Spectroscopy, (Academic Press, New York and London, 1972).
36. P. B. Fellgett, “Conclusions on multiplex methods,” Journal de Physique, Colloque C2, Supplément au n. 3–4 Tome 28, mars-avril 1967, pp.: 165–171, (1967).
37. P. B. Fellgett, “I. — les principes généraux des méthodes nouvelles en spectroscopie interférentielle A propos de la théorie du spectromètre interférentiel multiplex,” J. Phys. Radium 19(3), 187–191 (1958). [CrossRef]
38. P. B. Fellgett, “The nature and origin of multiplex Fourier spectrometry,” Notes Rec. R. Soc. 60(1), 91–93 (2006). [CrossRef]
39. F. D. Kahn, “The signal: noise ratio of a suggested spectral analyzer,” Astrophys. J. 129, 518–520 (1959). [CrossRef]
40. W. Schumann, and T. S. Lomheim, “Infrared hyperspectral imaging Fourier transform and dispersive spectrometers: comparison of signal-to-noise-based performance,” Imaging Spectrometry VII, San Diego, CA, USA, SPIE vol. 4480, (2001).