Suppression of the near-infrared OH night sky lines with fibre Bragg gratings – first results

For each frame the detector was used in a multiple-read mode (MRM) using an up-the-ramp sampling, with intervals of either 5, 15 or 30s and either 13 or 61 reads, see Table 2. This means that the detector is re-set, then read non-destructively at a fixed time interval for a specified number of times. This enables non-linearity in the detector response to be removed. Upon investigation we found significant non-linearity in the first few reads. For dark frames with the blank filters in, if one always observes in MRM then the first read after re-set is higher than the second read, and thereafter the counts are linear until about $20,000$ ADU (see Figure 2a). If the MRM exposure follows a double read mode observation (DRM; i.e. only two reads immediately prior to and after the exposure), then the first $\sim 5$ reads show a decline in the bias levels (see Figure 2b). This effect has been seen by one of us (AJH) in other Hawaii-1 PACE detectors. The largest observed drop in bias levels in the first few reads is $\approx 15$ ADU, which is insignificant under typical non-OH suppressed observations since the brightness of the night sky swamps such small variations. Thus for normal NIR observations a linear least-squares fit to the reads suffices to recover the incident count rate. However, in the case of OH suppressed observations this is no longer so, since the incident count rate can be very low for pixels between the OH lines. Therefore we have simply dropped the first 5 reads from our data cubes, before performing a linear least squares fit to the remaining data. The slope of this fit is then multiplied by the original number of reads minus one to recover the true incident count rate.

The detector non-linearity due to the filling of the pixel wells is mitigated by keeping the counts per pixel below 20,000 ADU. Thus in collapsing the data cubes as described above, we use a linear least squares fit. Thereafter we correct for any non-linearity following the procedure given on the IRIS2 web pages (http://www.aao.gov.au/AAO/iris2/iris2_linearity.html).

The GNOSIS background consists of dark current, instrument thermal background, telescope thermal background, and the night sky background (including OH emission, thermal emission, zodiacal scattered light, moonlight etc.). For our object observations we nod the telescope between object and sky, and thus all the background can be subtracted simultaneously using the sky observations. Pairs of observations are first subtracted, then averaged using the median if there is more than two.

For our sky observations this is not possible. In this case we subtract the dark current and the thermal background separately. The dark current is subtracted using the median of several dark frames of the same read mode and exposure time as the observations.

We have measured the thermal component from “cold frame” observations. A cold frame is an observation for which the GNOSIS fore-optics were removed from the Cassegrain focus and pointed directly into a dewar filled with liquid nitrogen. Thus the cold frame contains the GNOSIS fore-optics and IRIS2 thermal backgrounds and the detector dark current alone. The telescope thermal background should be much lower than the instrument background since the emissivity from the mirror should be less than $3$ per cent, and the emissivity from the Cassegrain hole is from a much smaller solid angle.

The average cold frames are shown in Figure 3. The counts have been calibrated to reproduce a $T=282$ K, 100 per cent emissivity black body, i.e. the expected emission from the slit block etc. at the output end of GNOSIS. This calibration provides an independent check of the efficiency of IRIS2 as discussed in section 4.1.

Obtaining cold frames is a significant overhead on observing time since the fore-optics must be removed and then re-fitted and re-aligned during the night when the optics are at the same temperature as for the sky frames. Thus in practice we fit a black-body spectrum to our observations by choosing regions of the continuum between the OH lines with $\lambda>1.7$$\mu$m. The best fitting model is subtracted from observations. We note that thermal emission is only approximately black-body (see Figure. 3). However, a more general function may inadvertently fit and subtract intrinsic features of the sky emission. Since we do not know the exact function of the thermal emission we revert to an approximate, but physically motivated black-body fit. An example black-body fit is shown in Figure 4. This is actually done after the residual background correction step detailed in  § 3.1.8 below.

Spectroscopic observations of the illuminated dome flat field screen were taken to provide a trace of the spectra across the detector. The centre of each of the 7 spectra is calculated at each spectral pixel. This is done by fitting the sum of seven Gaussians to the pixel counts in the spatial direction at each spectral pixel. The Gaussians for each fibre are assumed to have identical $\sigma$ and to be equally spaced. The fit yields the centre of each Gaussian, and this is stored in a look-up table, along with the $\sigma$ of the Gaussian fit. The separation between the spectra is 11 pixels, which is $\approx 5.5$ times the FWHM of the PSF, and therefore cross-talk between the spectra is negligible.

Object spectra are then extracted following the “Gaussian summation extraction by least squares” method of Sharp & Birchall (2010), i.e. the count for each spectral pixel, $c$ is given by,

where the sum ranges over each pixel, $i$, in the spatial direction between $\pm\frac{1}{2}$ the pitch between spectra, $D_{i}$ is the count at each pixel, $\sigma_{i}^{2}$ is the variance on the count, and $\phi_{i}$ is the value of the normalised Gaussian profile at that pixel. This method minimises the contribution due to noise in the extraction.

We next correct for variations in throughput between fibres. The fibre-to-fibre variation is measured from dispersed images of the illuminated dome flat field screen which are extracted in the same manner as the object spectra. The normalised fibre-to-fibre variation is the ratio of the total flux in each spectrum to the average of the total flux of all seven spectra. The extracted spectra are divided by the normalised variation.

The seven extracted spectra are combined by taking the sum of each spectral pixel. If a control fibre was used, this is excluded from the combination. This step neglects to take into account differences in the wavelength solution of each spectrum, which are less than 1.5 pixels over most of the detector. We have checked that not correcting for these differences does not introduce spurious signals by repeating the reduction of the sky spectrum for a single fibre only. We find that the single extracted spectrum is consistent with that of the summed spectrum.

The wavelength calibration is achieved using Xe arc lamp spectra. These spectra are extracted and combined as described above. The Xe lines are automatically detected and a cubic polynomial is fit to give the wavelength as a function of spectral pixel number. This solution is then applied to the object spectra. The wavelength calibration is accurate to $\approx 2$ Å, and is limited by the slight shift in the dispersion solution for each spectrum.

Hawaii-1 detectors are known to suffer from significant inter-quadrant cross-talk (Tinney et al. 2003¹¹1See also
http://www.eso.org/$\sim$gfinger/hawaii_1Kx1K/crosstalk_rock/crosstalk.pdf); if a bright object appears at pixel $\left\{x,y\right\}$, then there will also be a faint glow at pixel $\left\{x,{\rm mod}(y+512,1024)\right\}$ (where ${\rm mod}$ is the modulo operation), and along the entire row $x$. The cross-talk between quadrants does not affect GNOSIS observations, since all seven spectra are located on the lower two quadrants, but the cross talk along rows does.

The cross-talk along a row can be measured in the region of the spectrum at $\lambda<1.45$$\mu$m, since this region ought to be completely dark due to the spectroscopic blocking filter. This measurement may also include other residual background from systematic errors in the previous subtraction, due to changing background levels. We measure the mean count rate at $\lambda<1.45$$\mu$m and subtract it from the spectrum.

The efficiency of the GNOSIS system, including the atmosphere, the telescope and IRIS2, is measured from observations of A0V stars. These observations are reduced as described in the previous steps. We then take a model spectrum of Vega (Castelli & Kurucz 1994) and divide the observations by this to give a relative throughput of the system as a function of wavelength, including telluric features. Note that this will also include the averaged pixel-to-pixel variation in our extracted spectra, and hence includes the detector flat-field response. The relative throughput is normalised between 1.5 and 1.69 $\mu$m to give the instrument response function. The object spectra are then divided by the instrument response function.

The observations of the telluric standard stars described in the previous step can also be used to estimate the flux calibration. Since the magnitude of the A0V stars is already known, we can scale the model Vega spectrum to the appropriate brightness before dividing the observed spectrum to yield a flux calibrated correction. This step is not very accurate since scaling the Vega spectrum relies on knowing the aperture losses of the IFU at the time of observation, which in turn depends upon the seeing and acquisition, which are not readily measurable. The flux calibration is discussed further in § 4.1.2.

Due to the difficulty of measuring the seeing and the centring of our objects in the IFU we have computed the aperture losses for a seven element hexagonal array, assuming the seeing profile is Gaussian. The results are shown in Figure 5 as a function of seeing for several values of mis-alignment. Assuming that we can accurately align to within one element of the IFU, i.e. the offset is $\leq 0.2$ arcsec, and the seeing is 1.2 arcsec (extrapolated from measurements taken with the acquisition camera in the I band during the observations), then the typical aperture loss for a point source is $\leq 0.58$.

The throughput of GNOSIS was measured in the laboratory using a super-continuum source and a near-infrared camera (see Trinh et al.  in preparation for details). The total end-to-end throughput of GNOSIS is $\approx 36$ per cent and the throughputs of the individual components are given in Table 3 with the estimated errors.

The throughput of the AAT is approximately $0.94^{2}$ for the two reflections off the aluminium coated primary and secondary mirrors, which is estimated from a $0.98$ reflectivity of bare Al, combined with an extra $0.96$ throughput from the accumulation of dirt on the mirrors etc. This extra loss is compatible with that measured in the visible on the AAT.

We have estimated the throughput of IRIS2 in two ways. First we used the IRIS2 imaging exposure calculator to estimate the efficiency of an imaging observation, which for H is $\approx 29$ per cent. Combined with the throughput of the grism ($\approx 40$ per cent) this gives a total efficiency of $\approx 12$ per cent. Secondly we used the observations of the “cold frames” (see § 3.1.3) compared to a theoretical blackbody spectrum taking into account the area and solid angle of the IRIS2 slit-mask holes and assuming $T=282$ K as recorded, and 100 per cent emissivity, since most of the emission will come from the slit block and other mechanical parts surrounding the fibres due to the oversized slit mask holes. This also yielded an efficiency of 12 per cent.

The end-to-end system throughput is therefore expected to be, $\approx 3.6$ per cent, or $\approx 1.5$ per cent for a point source including typical aperture losses. We emphasise that this low throughput is not an intrinsic property of OH suppression systems, but is rather a consequence of retro-fitting an OH suppression unit to an existing spectrograph, which already has only 12 per cent throughput. The throughput of GNOSIS itself is 36 per cent and is limited mainly by the photonic lanterns. This will be discussed further in § 5.

We have also measured the end-to-end throughput as a function of wavelength from our observations of A0V stars. Examples are shown in Figure 6. The notches in the FBGs are visible as the significant narrow dips in the throughput. Note that in this graph the notches have been convolved with the spectrograph point spread function, so they appear shallower than the actual suppression depth. The measurement for HIP 104664 does not include the control fibre. The measured values are very sensitive to seeing and aperture losses as shown in Figure 5, but are within the errors of the anticipated throughput estimate in Table 3. In the case of HIP 109476 the measured throughput would require very low aperture losses due to exceptional seeing, suggesting some of the values in the table may be pessimistic. However, we assume the conservative values in Table 3 for the rest of the paper.

The OH lines between 1.5 – 1.7 $\mu$m are strongly suppressed. We have measured the suppression factor for 57 doublets; for the remaining 46 doublets we were unable to obtain a good fit to the unsuppressed line due to either line blending, intrinsic faintness or due to the filter cut-off (at $<1.5\ \mu$m). In most cases the measured strength of the suppressed line is poorly constrained because the flux is so low. Nevertheless, of the 57 doublets 78 per cent meet or exceed the target specifications.

Some quite bright lines are not suppressed by the FBGs. These are marked with black and green vertical lines in Figure 7. The green line is an O₂ emission line resulting from the a-X v=0 – 1 vibrational transition (Jeremy Bailey, private communication). This line is very bright in the day time and fades rapidly throughout twilight, but there is faint persistent emission throughout the night. It was not included in our FBG designs.

The lines marked in black are OH lines from the Q1(3.5) and Q1(4.5) rotational lines in the 3–1, 4–2 and 5–3 vibrational transitions. The FBGs have notches of $\approx 20$ dB for all these lines, printed at the wavelengths given by Rousselot et al. (2000). However, we measure a mean suppression factor of only $\approx 2.5$ dB. There may be an issue with the printed notch wavelengths of these particular transitions due to a larger spacing between the individual $\Lambda$ doublets than for other transitions (Pierre Rousselot, private communication). We show the wavelengths and gap size of the transitions in question as modelled by Rousselot et al. (2000) and as measured by Abrams et al. (1994) in Table 4. The mean wavelengths of the doublets are identical in each case, but the measured gap size is considerably larger; this is not a problem for the Q1(0.5), Q1(1.5) and Q1(2.5) transitions for which the $\Lambda$ doublets are much more closely spaced ($<100$ pm). However for the Q(3.5) and Q(4.5) transitions the doublet spacings derived from the Abrams et al. (1994) measurements are larger than the typical GNOSIS FBG notch widths, i.e. the individual lines fall either side of the notch. These lines account for 37 per cent of the doublets which did not meet the target suppression depth.

We have measured the level of suppression at each pixel by dividing the control spectrum by the suppressed spectrum (after first correcting the control spectrum for the fact that this is from a single fibre compared to six fibres for the suppressed spectrum), this is shown in Figure 8. The mean reduction per pixel between $1.5$ and $1.7$ $\mu$m is $\approx 17$, and the median is $\approx 1.6$. The reduction of the integrated background between $1.5$ and $1.7$ $\mu$m is $\approx 9$.

To compare the interline components in the suppressed and control spectra we first subtract the detector dark current and instrument thermal emission. However, these components are brighter than the interline component we are trying to measure, and the subtraction is inevitably noisy. Figure 9 shows a comparison of the electrons per second per pixel due to dark current, thermal emission and suppressed and unsuppressed OH lines. The statistical fractional error on the suppressed observations is shown in Figure 10, which is calculated from the mean and standard deviation of counts per pixel from15 separate half-hour exposures. The typical error on the interline component is $\approx 25$ per cent; however there may be unquantified systematic errors on these measurements.

Nevertheless, after dark and thermal subtraction, there is no discernible reduction in the background between the OH lines. This is unexpected; the models of Ellis & Bland-Hawthorn (2008) show that if the background (after dark subtraction) is composed of thermal emission, zodiacal scattered light and OH lines then the interline region should be dominated by light that originates as OH lines that is scattered by the spectrograph. Thus suppressing the OH lines should also suppress this light scattered from the OH lines. The fact that we do not see this suggests four possibilities: (i) there is residual instrumental emission which has not been properly removed in the data reduction, (ii) the IRIS2 spectrograph does not scatter significant amounts of OH light between the lines, (iii) the interline region is dominated by some source that Ellis & Bland-Hawthorn (2008) did not account for, (iv) the FBGs are not suppressing the OH lines as predicted, and hence the reduction is not as predicted.

We have addressed (i), the possibility of residual thermal emission, as fully as possible in our data reduction method (§ 3.1). However we are operating in the regime of very low count rates, such that systematic errors may become significant. The raw ADU per pixel in the interline component region is $\sim 10$ in a half hour exposure, which corresponds to  $\sim 45$ e^- per pixel, of which $\sim 18$ e^- per pixel are due to the sky and the detector dark current contributes $\sim 27$ e^- per pixel . That is we are detecting less that 1 e^- per pixel per minute, and we are dominated by detector noise. At this level systematics become very important. For example the effective read noise is $\approx 8$ e^- per pixel which alone gives $\sim 15$ per cent uncertainty in the interline component measurements. There are systematics in detector linearity such as discussed in section 3.1.1, and there may be other effects such as reciprocity failure (Biesiadzinski et al. 2011) which are very difficult to characterise.

We can discount (ii) since we have measured the scattering of IRIS2, and indeed it was this empirical scattering model from IRIS2 that was used in the models of Ellis & Bland-Hawthorn (2008). There is a small change in the PSF of IRIS2 due to a different slit and a different f-ratio at the input, however this will affect only the Gaussian core of the PSF, and then only slightly, and not at all the Lorentzian wings of the PSF.

Regarding (iii), we have searched for other molecular bands and lines of emission in the H band. We used the SpectralCalc.com facility to compute the emission line intensities from the HITRAN2008 models (Rothman et al. 2009). We included lines from H₂O, CO₂, O₃, N₂O, CO, CH₄, O₂, NO, SO₂, NO₂, NH₃, HNO₃, N₂, HF, HCl, HBr, HI, ClO, OCS, H₂CO, HOCl, HCN, CH₃Cl, H₂O₂, C₂H₂, C₂H₆, PH₃, COF₂, SF₆, H₂S, HCOOH, HO₂ and ClONO₂. The relative intensities were taken directly from the atmospheric paths of SpectralCalc.com assuming an observing altitude of 1.2 km, a target height of 600km, standard atmospheric conditions and zenith angle of 0 deg. The resulting emission is shown in Figure 11 where we have normalised the intensity to match roughly the interline component. There are hints that some of the structure in the interline component arises from certain molecular bands. The O₂ band at 1.58 $\mu$m and a CH₄ feature at $\approx 1.667\ \mu$m are most obvious, but there are possibly some other broad features due to CO₂, N₂O and C₂H₂. However the relative intensities of these features are not right, and neither is the overall shape of the combined emission from these molecules. Moreover, with the exception of O₂ and CH₄ there are no features corresponding to the obvious emission lines seen throughout.

Scattering from aerosols or dust in the upper atmosphere, and certain chemical reactions may also produce a nightglow continuum. For example the continuum produced in the reaction NO $+$ O $\to$ NO${}_{2}+$ h$\nu$, has been found to correlate well with the optical nightglow (Sternberg & Ingham 1972; Sternberg 1972), and it has been suggested (Content 1996)that this may increase in the NIR and is of the same order as the brightness of the interline component measured by Maihara et al. (1993). We have measured the interline component strength for individual observations as a function of airmass and find no trend, see Figure 12. This suggests that the interline component does not have an atmospheric origin, but we caution that the atmospheric emission could be temporally variable, as for the OH lines, which could mask any dependence on airmass. In summary there are individual features of the interline component due to other atmospheric molecular emission, but as yet we cannot explain the overall structure of the interline emission.

Finally we consider (iv), that there may still be unsuppressed OH emission dominating the interline component (e.g. the Q1(3.5) and Q1(4.5) transitions discussed in § 4.2.1). We have tested this and find a weak correlation between the OH line strength and the interline component strength for individual observations (recall that the OH line strength varies by about 10 per cent on the timescale of minutes Ramsay et al. 1992; Frey et al. 2000 and by a factor of $\approx 2$ throughout the night Shimazaki & Laird 1970), shown in Figure 13. The r² goodness of fit is 0.38, with a significance of $p<0.02$. Thus the OH line intensity does indeed affect the interline component. Nota bene, this is expected even with OH suppression, and is not an indication that the FBGs are not correctly suppressing the OH lines. Ellis & Bland-Hawthorn (2008) show that at a resolving power of $R=2000$ the scattering from residual suppressed OH lines, even though very faint, will still dominate the interline region even after OH suppression. In fact it is the relative weakness of the correlation between interline component and OH line strength that must be explained, especially when coupled with the lack of dependence on airmass in Figure 12. This will at least in part be due to uncertainty in the measurements. The signal to noise on the interline component is poor, and the standard deviation of the residuals to the best fit shown in Figure 13 is 160 photons s^-1 m^-2 arcsec^-2 $\mu$m^-1, which gives an indication as to the error on the interline component measurements. However, the weakness of the correlation may also be further indication that there is still some unidentified source of interline emission.

In summary, we feel confident that we can rule out (ii), i.e. we believe we understand the scattering function, but cannot rule out any of the other options. Instrumental artifacts, unknown atmospheric components and inaccurate modelling of the OH lines may all contribute to the interline component at some level.