JSTOR

1. INTRODUCTION

Since ancient times, astronomers have sought to draw maps of the sky, similar to their cartographer colleagues’ drive to produce ever more accurate maps of the Earth. However, it was not until the invention of photography that faint, large-scale structures of nebulae and dust lanes could be faithfully reproduced. Since the late nineteenth century, several visible-light, wide-angle photographic surveys of the sky have been completed, such as E. E. Barnard’s Atlas of Selected Regions of the Milky Way (Barnard et al. 1927); the National Geographic Society-Palomar Observatory Sky Survey of the northern sky (Minkowski & Abell 1963) and subsequently the equivalent ESO/SERC atlas of the southern sky; and more recently, the Sloan Digital Sky Survey (York et al. 2000). In addition, a Hα narrowband optical survey was compiled (Finkbeiner 2003). Very recently (after the original manuscript for the current article was submitted), the European Southern Observatory (ESO) released a Milky Way panorama image photographed by S. Brunier as part of their Gigagalaxy Zoom project.1

In 2000, the author published an all-sky color image taken on photographic 35 mm color film with a focal length of f = 28 mm. A total of 51 images were computer processed into a mosaic measuring 14400 × 7200 pixels, equivalent to an image scale of 90^″pixel^-1 (di Cicco 1999; Kizer Whitt & Mellinger 2001). This picture is now widely used for educational purposes in planetariums, exhibitions, and textbooks; it formed the basis for a photographic star atlas (Mellinger & Hoffmann 2005), and illustrates the location of star-forming regions in the Milky Way (Mellinger 2008). With recent advances in digital projection technology, demand grew for both a higher spatial resolution and a higher dynamic range. The article describes the assembly of a 648 megapixel mosaic from more than 3000 individual CCD images.

2. DATA ACQUISITION

2.1. Hardware

2.1.1. CCD Camera Wide-field imaging equipment based on commercial camera lenses has been successfully used in the past for, e.g., emission-line surveys (Gaustad et al. 2001) and measurements of the night-sky brightness (Shamir & Nemiroff 2005; Duriscoe et al. 2007). After exploration of several camera systems on the market, an SBIG STL-11000 CCD camera was chosen due to its

1.	Moderate cost,
2.	Large sensor size (36 mm × 24 mm),
3.	Antiblooming gate (eliminating the need for software deblooming), and
4.	Compact construction (low back focus, allowing the use of 35 mm camera lenses).

The camera uses a Kodak KAI-11002 chip, a 4008 (H) × 2672 (V) progressive scan interline CCD image sensor with an antiblooming gate (Eastman Kodak Company 2006). It uses microlenses to enhance its quantum efficiency by directing light to the active pixel areas. Drawbacks of this design include nonlinearity and angular quantum-efficiency variations, and are addressed in §§ 3.1 and 3.2.

2.1.2. Lens The camera was fitted with a Minolta MD 1.4/50 mm lens originally used on a 35 mm format film-based single-lens reflex camera. To improve the image quality, the lens was stopped down to f/4 for all exposures.

2.1.3. Filters As the principal purpose of the panorama image is to create a visual representation of colors in the sky, an RGB filter set (Astronomik Type 2c) was used instead of the Johnson UBVR filters common in photometry. The transmission curves are shown in Figure 1.

(23 KB)

Fig. 1.— Transmission curves of the Astronomik RGB filters.

Open New Window

2.2. Field Selection

The KAI-11002 chip has an active area of 36 × 24 mm², resulting in a field of view of approximately 40° × 27° at a focal length of f = 50 mm. Due to the location of the mounting thread on the camera, a “portrait” orientation of the camera was chosen, with the long side of the chip being parallel to circles of constant right ascension. To ensure a seamless mosaic, adjacent fields had to overlap each other by at least 15% of the respective offset in right ascension or declination. With these constraints, a coverage pattern (Fig. 2) was developed by way of trial and error. The pattern consists of five rings of frames centered at declinations of +62°, +32°, 0°, -32°, and -62°, in addition to two fields centered at the celestial poles (± 90°). The number of fields and offsets in right ascension are listed in Table 1.

TABLE 1 Declination of Field Centers and R.A. Offsets

Open New Window

(59 KB)

Fig. 2.— Northern (left) and southern (right) hemisphere coverage maps of the 70 fields.

Open New Window

2.4. Sites

The fields centered at declinations -90°, -62°, -32°, and 0° were taken in 2007 October and 2008 March/April from dark-sky locations in South Africa (Tables 2 and 3), while the northern hemisphere images (declinations +32°, +62°, and +90°) were taken between 2008 December and 2009 August from dark-sky locations in Texas and Michigan in the US.

TABLE 2 Observation Sites

Open New Window

TABLE 3 Field List

Open New Window

3. DATA PROCESSING

3.1. Nonlinearity Correction

The use of a CCD sensor with an antiblooming gate often comes at the expense of linearity (Duriscoe et al. 2007). To test the sensor linearity, a uniform white target illuminated by a light-emitting diode (LED) operated in constant-current mode was photographed with a series of exposure times ranging from 0.02 to 30 s. To check for variations in the LED intensity, two series of exposures (one from short to long exposure times, the other from long to short) was taken. After dark-frame subtraction, the flux in a field measuring 200 × 200 pixels at the image center was measured using the IRAF imstat task. The test revealed a significant deviation from a linear response for a signal as low as 15000 ADU s pixel^-1, well below the full-well depth of approximately 60000 ADU s pixel^-1 (Fig. 3). However, the observed signal as a function of illumination could be well represented by the quadratic equation where E is the exposure (product of light intensity and exposure time) and S is the measured signal. Based on these data, a nonlinearity correction was developed as follows: Solving equation (1) for E yields The corrected (linearized) signal is then whereFrom a least-squares fit, α was determined as -7.5334 × 10^-7 ADU^-1. This function was applied to all raw images, including the flat frames (but not the dark frames, due to their low ADU values).

(19 KB)

Fig. 3.— Nonlinearity of the STL-11000 camera obtained in laboratory tests with a constant-current LED light source. The solid line shows the least-squares fit according to eq. (1).

Open New Window

3.2. Preprocessing

The standard preprocessing steps (dark-frame subtraction, masking of bad pixels, division by flat frame) were applied to all images. Due to its microlenses, the angular quantum efficiency of the CCD chip depends upon the angle of incidence. This angular profile in turn is a function of the wavelength (Eastman Kodak Company 2006). Hence, flat frames taken through the R, G, and B filters differ significantly (Fig. 4).

(45 KB)

Fig. 4.— Ratio of R and G flat fields.

Open New Window

Depending on the spectral distribution of the sky background, this wavelength-dependent sensitivity of the camera system may lead to residual gradients, even with proper flat-frame calibration. However, these gradients were later eliminated by the background-subtraction procedure detailed in § 3.4.3.

Next, the images were upsampled by a factor of 3. Although it increases the file size by 9×, this step was found to be necessary in order to minimize the loss of resolution in subsequent resampling with bilinear interpolation. Higher-order interpolation methods, such as bicubic or Lanczos resampling, could not be used as they introduce ringing artifacts in undersampled images.

3.4. Photometric Calibration

A major problem was the photometric calibration of the individual fields. A straight combination without photometric corrections yields a patchy-looking mosaic, as shown in Figure 5. Clearly, the sky background level varies significantly between fields, and some fields exhibit strong intrafield background gradients. Moreover, differences in atmospheric transparency may cause the recorded flux I of a star to vary between two frames.

(63 KB)

Fig. 5.— R channel mosaic, combined without photometric calibration and background subtraction. There are strong variations in the background level, as well as intrafield gradients.

Open New Window

3.4.1. Contributions to the Sky Background The diffuse night-sky background has a variety of sources (see Leinert et al. 1998 for an in-depth discussion):

1.	Light pollution (ILP). Although most images were taken from dark-sky sites (class 1 on Bortle’s scale (Bortle 2001)), time constraints dictated that some fields were taken from sites with slight artificial light pollution (the Michigan sites have Bortle class 2 or 3). In addition, again due to time constraints, some images were taken in the evening sky before the end of astronomical twilight, or in the morning sky past the beginning of astronomical twilight.
2.	Airglow (IA). In the absence of light pollution, airglow is the largest contributor to the night sky brightness (Leinert et al. 1998). In the visible spectrum, the main sources are atmospheric O2 (for wavelengths λ < 430 nm) and OH (for λ > 520 nm). Air glow is highly variable both in spatial distribution and in time.
3.	Zodiacal light (IZL). Zodiacal light is sunlight scattered by interplanetary dust particles. Near the ecliptic poles, its intensity in the visible spectrum is typically 2–4 times weaker than airglow, but it gains significantly as the field of view gets closer to the sun. Due to its spatial confinement to the ecliptic region, it can cause strong background gradients in wide-field images.
4.	Tropospheric scattering (Isca). Starlight as well as airglow and zodiacal light are scattered in the troposphere. High clouds as well as high humidity at ground level can enhance scattering and lead to halos around bright stars.
5.	Integrated starlight (IISL). Another contribution to the sky brightness is the light from unresolved stars. At visible wavelengths, it is dominated by main-sequence stars (Leinert et al. 1998).
6.	Diffuse galactic light (IDGL). Diffuse galactic light (DGL) is produced when starlight is scattered by dust grains in interstellar space.
7.	Extragalactic background light (IEBL). Light from galaxies that are not individually detected.

Many of these contributions to the night-sky brightness are undesirable in an all-sky mosaic image. Light pollution, airglow, zodiacal light, and tropospheric scattering vary between individual fields, as well as within a field, and thus should be removed in the calibration process. However, the contribution from unresolved stars and the diffuse galactic light are features of the Milky Way and should be preserved, together with any extragalactic background. This creates a major calibration problem, as we need to decide on how much of the local background is caused by unwanted sources (1–4 in list) or galactic and extragalactic contributions to be retained (5–7 in list). This problem has plagued a number of ground-based studies of the galactic background (Leinert et al. 1998). However, these effects can be separated using space-borne visual photometry data from the Imaging Photopolarimeters (IPPs) onboard the Pioneer 10 and 11 space probes. Their data were obviously not contaminated by terrestrial sources, and the zodiacal light contribution became negligible for distances above 3.3 AU (Gordon 1997; Gordon et al. 1998).

The photometric calibration of the all-sky image was performed in two steps:

3.4.2. Standard Star Processing For each field and color channel, a list of several hundred reference stars (that included their equatorial coordinates, fluxes S (in ADU), and local backgrounds β) was generated using Source Extractor and matched against the Yale Bright Star (YBS) Catalog.2 Typically, 100…120 matches were found. The reference star magnitude V and the measured flux are related via where Q is the instrumental zero point (including effects of atmospheric extinction) and (B - V) is the color index from the YBS Catalog. Q and ϵ were then determined via a least-squares fit. A typical plot of calibrated V magnitudes (calculated according to eq. [5]) versus YBS catalog magnitudes is shown in Figure 6. As expected, the spectral correction parameter ϵ was close to -1 for the B channel frames and close to 0 for the G channel, while values of approximately +0.3 were typical for the R channel. The ADU units were then converted to S₁₀(V) units commonly used for measuring surface brightness (Sparrow & Weinberg 1976). One S₁₀(V) unit represents the sky brightness resulting from one tenth-magnitude star per square degree. Thus, where κ is the image scale in degrees per pixel.

(14 KB)

Fig. 6.— Calibrated V magnitudes (calculated the measured flux via to eq. (5)) vs. YBS catalog magnitudes.

Open New Window

3.4.3. Background Subtraction According to the discussion at the start of § 3.4, the total light intensity seen by a CCD pixel is where S is the flux due to point sources or resolved diffuse objects, I_bad = I_LP + I_A + I_ZL + I_sca is the unwanted background contribution from terrestrial and solar system sources, and I_good = I_ISL + I_DGL + I_EBL is the galactic and extragalactic background to be retained in the image. Since I_good is known from the Pioneer 10/11 data (Gordon et al. 1998), the spatial distribution of I_bad can be modeled by fitting a surface to (I - I_good).

Unfortunately, the high-resolution background data by Gordon et al. (1998) do not cover the entire sky; they excluded the data contaminated with scattered sunlight, taken within 70° of the Sun for Pioneer 10 and within 45° for Pioneer 11, respectively. Leinert et al. (1998) provide some additional data points on a 10° × 10° grid; the other missing data were interpolated using a bicubic spline function. Some local median filtering was applied in order to remove small-scale artifacts in the data.

With the Pioneer data providing I_good, the “bad” background was estimated by applying a strong median filter to (I - I_good) = (S + I_bad). Subtracting I_bad from I then yields the desired sum of S + I_good. This approach also eliminates any residual gradient from the wavelength-dependent angular response of the camera system (cf. § 3.2).

4. RESULTS

The above procedure created a 360° × 180° panorama image of the entire sky at a scale of 36″pixel^-1, equivalent to a size of 36000 × 18000 pixels. With each pixel represented by a 32 bit (single precision) floating-point number, each color channel is approximately 2.5 gigabytes in size. The faintest visible stars are approximately 14 mag, while only stars brighter than +0.7 mag are saturated. (This saturation could have been avoided by reducing the shortest exposure time below 0.5 s, but this would have reduced the number of reference stars suitable for astrometric calibration.) Hence, the dynamic range is approximately 18 bits.

A low-resolution color image of the panorama is shown in Figure 7. A zoomable Mercator projection can be seen on the author’s web site.4 The image provides a detailed view of large-scale galactic structures, such as a halo of unresolved stars extending to b ≤  ± 25°, and numerous dark clouds. Note the strong reddening of starlight in the Ophiuchus/Scorpius region of the Milky Way. In the earlier panorama (Mellinger & Hoffmann 2005), many of these details are missing due to the rather aggressive flat-fielding employed at the time. In a historical contect, the panorama image may be compared to the hand-drawn maps of Gaposchkin (1957) and the first all-sky image in Aitoff projection by Lundmark,5 as well as early photometric studies of the Milky Way (Pannekoek et al. 1949).

(120 KB)

Fig. 7.— RGB mosaic, combined with photometric calibration and background subtraction. Top: entire sky in Hammer-Aitoff projection. Middle: enlarged view of the central Milky Way. Bottom: full-scale view of the Antares/ρ Oph region.

Open New Window

Besides galactic stars and nebulae, and some brighter extragalactic objects (most notably, the Magellanic Clouds and the Andromeda galaxy), the panorama image also shows a number of solar system objects. In most cases, the bright planets could be avoided by scheduling the exposures for times when no planet was near the field of view. No attempt was made to remove or avoid solar system objects invisible to the naked eye.

Several of the raw images taken from South Africa in 2008 April exhibit trails of geosynchronous satellites. At that time (2–3 weeks after the southern hemisphere fall equinox), sunlight glinting off reflective surfaces is directed toward the southern hemisphere, causing the satellite to brighten by several magnitudes. As the camera was tracking the stars, each satellite left a 1° trail on the 240 s exposures. Median combining the raw frames eliminated most of the satellite trails; however, fields #34 and 36 had to be retaken as the blending of multiple trails caused them to be quite apparent even in the combined images.

A comparison with the very recently released ESO Gigagalaxy Zoom Milky Way panorama6 shows that both images were photographed at the same focal length of f = 50 mm. Brunier made his panorama with a Nikon D3 camera which has a noncooled 36 × 23.9 mm² sensor with 4256 × 2832 pixels and a color filter array (CFA). Its pixels are about 6% smaller than those of the STL-11000; however, the demosaicing process required for CFA sensors (Gunturk et al. 2005) would lead to some loss of spatial resolution. No information on the dynamic range and photometric calibration of the ESO image is currently available, although the use of a single exposure time (6 minutes) suggests a dynamic range significantly lower than the 18 bit range of the image described in this article.

5. APPLICATIONS

First and foremost, the all-sky mosaic image is expected to be a valuable educational tool. Modern digital full-dome planetarium projection systems offer 12 bits of resolution per color channel, with 16 bits not too far down the road. With an 18 bit dynamic range, it will be possible to blend over from a view showing only the brightest stars (simulating the view in a light-polluted environment) to a full display of the faint star clouds and nebulae in our Milky Way. The resolution of 36″pixel^-1 ensures crisp star images that will be limited by the resolution of the projection equipment and the observer’s eye, rather than the image data. As the size of the image (7.5 gigabytes for a 32 bit floating-point representation) may exceed the capabilities of some personal computers, an OpenEXR version is in preparation. This file format uses 16 bit (half-precision) floating-point numbers for storing the image data,7 thus reducing the file size by a factor of 2.

In addition to its use in education, the image is also of scientific value, as it supplements earlier work on the surface photometry of the Milky Way, such as Hoffmann et al. (1998). As an example, studies of the night sky background (see, e.g., Duriscoe et al. 2007) need to discern between man-made light pollution and natural contributions to the sky brightness (a problem very similar to the one encountered in the present work), and would thus benefit from photometric data on the Milky Way brightness.