Publications of the Astronomical Society of the Pacific, 114:144–152, 2002 February

䉷 2002. The Astronomical Society of the Pacific. All rights reserved. Printed in U.S.A.

Drizzle: A Method for the Linear Reconstruction of Undersampled Images

A. S. Fruchter1 and R. N. Hook2
Received 2001 February 22; accepted 2001 October 10

ABSTRACT. We have developed a method for the linear reconstruction of an image from undersampled,
dithered data. The algorithm, known as Variable-Pixel Linear Reconstruction, or informally as “Drizzle,” preserves
photometry and resolution, can weight input images according to the statistical significance of each pixel, and
removes the effects of geometric distortion on both image shape and photometry. This paper presents the method
and its implementation. The photometric and astrometric accuracy and image fidelity of the algorithm as well
as the noise characteristics of output images are discussed. In addition, we describe the use of drizzling to combine
dithered images in the presence of cosmic rays.

1. INTRODUCTION shown for the HST and WFPC2 optics by the upper right panel
Undersampled images are common in astronomy because in Figure 1. Pixelated detectors then again convolve this image
instrument designers are frequently forced to choose between with the response function of the electronic pixel E; thus,
properly sampling a small field of view and undersampling a Id p T 丢 O 丢 E. The detected image can be thought of as this
larger field. Nowhere is this problem more acute than on the continuous convolved image sampled at the center of each
Hubble Space Telescope (HST), whose corrected optics now physical pixel. Thus, a shift in the position of the detector
provide superb resolution; however, the detectors on HST are (known as a “dither”) can be thought of as producing offset
able to take full advantage of the full resolving power of the samples from the same convolved image. Although pixels are
telescope only over a limited field of view. For instance, the typically square on the detector, their response may be non-
primary optical imaging camera on the HST, the Wide Field uniform and, indeed, may, because of the scattering of light
and Planetary Camera 2 (WFPC2; Trauger et al. 1994), is com- and charge carriers, effectively extend beyond the physical
posed of four separate 800 # 800 pixel CCD cameras, one of pixel boundaries. This is the case in WFPC2. By contrast, in
which, the planetary camera (PC), has a scale of 0⬙. 046 pixel⫺1, the NICMOS detectors (Storrs et al. 1999; Lauer 1999b), the
while the other three, arranged in a chevron around the PC, electronic pixel is effectively smaller than the physical pixel.
have a scale of 0⬙. 097 pixel⫺1. These latter three cameras, re- Fortunately, much of the information lost to undersampling
ferred to as the wide-field cameras (WFs), are currently the can be restored. In the lower right panel of Figure 1 we display
primary workhorses for deep imaging surveys on HST. How- an image made using one of the family of techniques we refer
ever, these cameras greatly undersample the HST image. The to as “linear reconstruction.” The most commonly used of these
width of a WF pixel equals the FWHM of the optics in the techniques are shift-and-add and interlacing. In interlacing, the
near-infrared and greatly exceeds it in the blue. In contrast, a pixels from the independent images are placed in alternating
well-sampled detector would have ⲏ2.5 pixels across the pixels on the output image according to the alignment of the
FWHM. Other HST cameras such as NICMOS, STIS, and the pixel centers in the original images. The image in the lower
future Advanced Camera for Surveys (ACS) also suffer from right corner of Figure 1 has been restored by interlacing dith-
undersampling to varying degrees. The effect of undersampling ered images. However, because of the occasional small posi-
on WF images is illustrated by the “Great Eye Chart in the tioning errors of the telescope and the nonuniform shifts in
Sky” in Figure 1. Further examples showing astronomical tar- pixel space caused by the geometric distortion of the optics,
gets are given in § 8. true interlacing of images is often infeasible. In the other stan-
When the true distribution of light on the sky T is observed dard linear reconstruction technique, shift-and-add, a pixel is
by a telescope, it is convolved by the point-spread function (PSF) shifted to the appropriate location and then added onto a sub-
of the optics O to produce an observed image, IO p T 丢 O, sampled image.
where 丢 represents the convolution operator. This effect is Shift-and-add can easily handle arbitrary dither positions,
but it convolves the image yet again with the original pixel,
1
compounding the blurring of the image and the correlation of
Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD
21218; [email protected].
the noise. In this case, two further convolutions are involved.
2
Space Telescope European Coordinating Facility, Karl-Schwarzschild- The image is convolved with the physical pixel P, as this pixel
Strasse 2, D-85748 Garching, Germany; [email protected]. is mathematically shifted over and added to the final image.

144
 DRIZZLE: RECONSTRUCTING UNDERSAMPLED IMAGES 145

Fig. 1.—Top left: “True image,” i.e., the image one would see with an infinitely large telescope. Top right: Image after convolution with the optics of HST
and the WFPC2 camera, the primary wide-field imaging instrument presently installed on the HST. Bottom left: Image after sampling by the WFPC2 CCD. Bottom
right: Linear reconstruction of dithered CCD images.

In addition, when many images with different pointings are The last convolution rarely produces a significant effect, how-
added together on the final output grid, there is also a con- ever, because the output grid is usually considerably finer than
volution by the pixel of the final output grid G. This produces the detector pixel grid, and convolutions add roughly as a sum
a final image: of squares (the summation is exact if the convolving functions
are Gaussians).
I p T 丢 O 丢 E 丢 P 丢 G. (1) The importance of avoiding convolutions by the detector

2002 PASP, 114:144–152

146 FRUCHTER & HOOK

pixel is emphasized by comparing the upper and lower right- from each input image. One must therefore choose a drop size
hand images in Figure 1. The deterioration in image quality that is small enough to avoid degrading the image but large
between these images is due entirely to the single convolution enough so that after all images are drizzled the coverage is
of the image by the WF pixel. The interlaced image in the reasonably uniform.
lower right panel has had the sampled values from all of the The drop size is controlled by a user-adjustable parameter
input images directly placed in the appropriate output pixels called pixfrac, which is simply the ratio of the linear size of
without further convolution by either P or G. the drop to the input pixel (before any adjustment due to the
Here we present a new method, Drizzle, which was originally geometric distortion of the camera). Thus, interlacing is equiv-
developed for combining the dithered images of the Hubble alent to Drizzle in the limit of pixfrac r 0.0, while shift-and-
Deep Field–North (HDF-N; Williams et al. 1996) and has since add is equivalent to pixfrac p 1.0. The degree of subsampling
been widely used for the combination of dithered images from of the output is controlled by the user through the scale pa-
both HST’s cameras and those on other telescopes. Drizzle has rameter s, which is the ratio of the linear size of an output pixel
the versatility of shift-and-add yet largely maintains the reso- to an input pixel.
lution and independent noise statistics of interlacing. While When a pixel (xi , yi ) from an input image i with data value
other methods (cf. Lauer 1999a) have been suggested for the dx i yi and user-defined weight wx i yi is added to an output image
linear combination of dithered images, Drizzle has the advan- pixel (xo , yo ) with value Ixo yo, weight Wxo yo, and fractional pixel
tage of being able to handle images with essentially arbitrary overlap 0 ! ax i yi xo yo ! 1, the resulting values and weights of that
shifts, rotations, and geometric distortion and, when given input same pixel, I
and W
, are
images with proper associated weight maps, creates an optimal
statistically summed image. Drizzle also naturally handles im- Wx
o yo p ax i yi xo yowx i yi ⫹ Wxo yo , (2)
ages with “missing” data, due, for instance, to corruption by
cosmic rays or detector defects. dx i yiax i yi xo yowx i yis 2 ⫹ Ixo yoWxo yo
Ix
o yo p , (3)
The reader should note that Drizzle does not attempt to im- Wx
o yo
prove upon the final image resolution by enhancing the high-
frequency components of the image that have been suppressed where a factor of s2 is introduced to conserve surface intensity
by either the optics or the detector. While such procedures, and where i and o are used to distinguish the input and output
which we refer to as “image restoration” (in contrast to “image pixel indices. In practice, Drizzle applies this iterative proce-
reconstruction”), are frequently very valuable (see Hanisch & dure to the input data, pixel by pixel, image by image. Thus,
White 1994 for a review), they invariably trade signal-to-noise after each input image is processed, there is a usable output
ratio for enhanced resolution. Drizzle, on the other hand, was image and weight, I and W.
developed specifically to provide a flexible and general method The final output images, after all inputs have been processed,
of image combination that produces high-resolution results can be written as
without sacrificing the final signal-to-noise ratio.
Wxo yo p ax i yi xo yowx i yi, (4)
2. THE METHOD dx i yiax i yi xo yowx i yis 2
Ixo yo p , (5)
Although the effect of Drizzle on the quality of the image Wxo yo
can be profound, the algorithm is conceptually straightforward.
Pixels in the original input images are mapped into pixels in where for these equations, (4) and (5), we use the Einstein
the subsampled output image, taking into account shifts and convention of summation over repeated indices and where the
rotations between images and the optical distortion of the cam- input indices xi and yi extend over all input images. It is worth
era. However, in order to avoid reconvolving the image with noting that in nearly all cases ax i yi xo yo p 0 , since very few input
the large pixel “footprint” of the camera, we allow the user to pixels overlap a given output pixel.
shrink the pixel before it is averaged into the output image, as This algorithm has a number of advantages over the more
shown in Figure 2. standard linear reconstruction methods presently used. It pre-
The new shrunken pixels, or “drops,” rain down upon the serves both absolute surface and point-source photometry
subsampled output. In the case of the HDF-N WFPC2 imaging, (though see § 5 for a more detailed discussion of point-source
the drops were given linear dimensions one-half those of the photometry). Therefore, flux can be measured using an aperture
input pixel—slightly larger than the dimensions of the output whose size is independent of position on the chip. And because
pixels. The value of an input pixel is averaged into an output the method anticipates that a given output pixel may receive
pixel with a weight proportional to the area of overlap between no information from a given input pixel, missing data (due for
the “drop” and the output pixel. Note that if the drop size is instance to cosmic rays or detector defects) do not cause a
sufficiently small, not all output pixels have data added to them substantial problem, so long as there are enough dithered im-

2002 PASP, 114:144–152

 DRIZZLE: RECONSTRUCTING UNDERSAMPLED IMAGES 147

Geometrical
Transformation

Output Fine
Pixel Grid
Original Coarse
Pixel Grid

Fig. 2.—Schematic representation of Drizzle. The input pixel grid (left) is mapped onto a finer output grid (right), taking into account shift, rotation, and
geometric distortion. The user is allowed to “shrink” the input pixels to smaller pixels, which we refer to as drops ( faint inner squares). A given input image
only affects output image pixels under drops. In this particular case, the central output pixel receives no information from the input image.

ages to fill in the gaps caused by these zero-weight input pixels. convolutions with both p and G effectively drop away. None-
Finally, the linear weighting scheme is statistically optimum theless, the small overlap between adjacent drops fills in miss-
when inverse variance maps are used as the input weights. ing data.
Drizzle replaces the convolution by P in equation (1) with
a convolution with p, the pixfrac. Because this kernel is usually
3. COSMIC-RAY DETECTION
smaller than the full pixel and as noted earlier convolutions
add as the sum of squares, the effect of this replacement is Few HST observing proposals have sufficient time to take
often quite significant. Furthermore, when the dithered posi- a number of exposures at each of several dither positions.
tions of the input images map directly onto the centers of the Therefore, if dithering is to be of widespread use, one must be
output grid, and pixfrac and scale are chosen so that p is only able to remove cosmic rays from data where few, if any, images
slightly greater than s, one obtains the full advantages of in- are taken at the same position on the sky. We have therefore
terlacing: because the power in an output pixel is almost entirely adapted Drizzle to the removal of cosmic rays. Because the
determined by input pixels centered on that output pixel, the techniques involved in cosmic-ray removal are also valuable

2002 PASP, 114:144–152

148 FRUCHTER & HOOK

Fig. 3.—Left: Region of one of 12 2400 s archival images taken with the F814W wide near-infrared filter on WFPC2. Numerous cosmic rays are visible. Right:
Drizzled combination of the 12 images, no two of which shared a dither position.

in characterizing the image fidelity of Drizzle, we will discuss 1993) scripts that are now available along with Drizzle in the
them first. Dither package of STSDAS. In addition to demonstrating how
Here then is a short description of the method we use for effectively cosmic rays can be removed from singly dithered
the removal of cosmic rays: images (i.e., images that share no common pointing), this image
also displays the degree to which linear reconstruction can
1. Drizzle each image onto a separate subsampled output improve the detail of an image. In the Drizzled image the object
image using pixfrac p 1.0. to the upper right clearly has a double nucleus (or a single
2. Take the median of the resulting aligned drizzled images. nucleus with a dust lane through it), but in the original image
This provides a first estimate of an image free of cosmic rays. the object appears unresolved.
3. Map the median image back to the input plane of each
of the individual images, taking into account the image shifts
and geometric distortion. This can done by interpolating the 4. IMAGE FIDELITY
values of the median image using a program we have named
“Blot.” The drizzling algorithm was designed to obtain optimal
4. Take the spatial derivative of each of the blotted output signal-to-noise ratios on faint objects while preserving image
images. This derivative image is used in the next step to es- resolution. These goals are unfortunately not fully compatible.
timate the degree to which errors in the computed image shift As noted earlier, nonlinear image restoration procedures that
or the blurring effect of taking the median could have distorted attempt to remove the blurring due to the PSF and the pixel
the value of the blotted estimate. response function through enhancing the high frequencies in
5. Compare each original image with the corresponding blot- the image, such as the Richardson-Lucy (Richardson 1972;
ted image. Where the difference is larger than can be explained Lucy1974; Lucy & Hook 1992) and maximum entropy methods
by noise statistics, the flattening effect of taking the median, (Gull & Daniell 1978; Weir & Djorgovski 1990), directly
or an error in the shift, the suspect pixel is masked. exchange signal-to-noise ratio for resolution. In the drizzling
6. Repeat the previous step on pixels adjacent to pixels al- algorithm no compromises on signal-to-noise ratio have been
ready masked, using a more stringent comparison criterion. made; the weight of an input pixel in the final output image
7. Finally, drizzle the input images onto a single output image is entirely independent of its position on the chip. Therefore,
using the pixel masks created in the previous steps. For this if the dithered images do not uniformly sample the field, the
final combination a smaller pixfrac than in step 1 will usually “center of light” in an output pixel may be offset from the
be used in order to maximize the resolution of the final image. center of the pixel, and that offset may vary between adjacent
pixels. Dithering offsets combined with geometric distortion
generally produce a sampling pattern that varies across the field.
Figure 3 shows the result of applying this method to data The output PSFs produced by the combination of such irreg-
originally taken by Cowie and colleagues (Cowie, Hu, & Son- ularly dithered data sets may, on occasion, show variations
gaila 1995). The reduction was done using a set of IRAF (Tody about the true PSF. Fortunately, this does not noticeably affect

2002 PASP, 114:144–152

 DRIZZLE: RECONSTRUCTING UNDERSAMPLED IMAGES 149

Fig. 4.—Example of the effect of geometric distortion on photometry. The pixels in the two images represent the photometric values of a 19 # 19 grid of stars on
the WF chip. Each pixel represents the photometric value of a star; the pixels are not images of the stars. The left panel shows the apparent brightness of the stars (all
of equal intrinsic brightness) as they would appear in a flat-fielded WF image. Because pixels near the edge of the chip are up to 4% smaller on the sky than pixels
near the center, the flat field has artificially brightened the stars near the edges and the corners. The right panel shows the photometric values of these stars after
drizzling, including the use of representative cosmic-ray masks (see Fig. 3). The standard deviation of the corrected stellar magnitudes is ≤0.015 mag.

aperture photometry performed with typical aperture sizes. In images in the corners of these images appear up to ∼4% brighter
practice, the variability appears larger in WFPC2 data than we in the corners of the images than near the center. These images
would predict based on our simulations. Examination of more were then drizzled with a scale p 0.5 and pixfrac p 0.6. Ap-
recent dithered stellar fields leads us to suspect that this excess erture photometry on the 19 # 19 grid after drizzling reveals
variability results from a problem with the original data, pos- that the effect of geometric distortion on the photometry has
sibly caused by charge transfer errors in the CCD (Whitmore been dramatically reduced: the rms photometric variation in
& Heyer 1997). the drizzled image is 0.004 mag. Of course this is not the final
photometric error of a drizzled image (which will depend on
5. PHOTOMETRY the quality of the input images) but only the additional error
Camera optics generally introduce geometric distortion of that the use of Drizzle would add under these rather optimal
images. In the case of the WFPC2, pixels at the corner of each circumstances.
CCD subtend less area on the sky than those near the center. In practice, users may not have four relatively well interlaced
This effect will be even more pronounced in the case of the images but rather a number of almost random dithers, and each
ACS. However, after application of the flat field, a source of dithered image may suffer from cosmic-ray hits. Therefore, in
uniform surface brightness on the sky produces uniform counts a separate simulation, we have used the shifts actually obtained
across the CCD. Therefore, point sources near the corners of in the WF2 F814W images of the HDF-N as an example of
the chip are artificially brightened compared to those in the the nearly random subpixel phase that large dithers may pro-
center. By scaling the weights of the input pixels by their areal duce on HST. In addition, we have associated with each image
overlap with the output pixel and by moving input points to a mask corresponding to cosmic-ray hits from one of the deep
their corrected geometric positions, Drizzle largely removes HST WF images used in creating Figure 3. When these sim-
this effect. In the case of pixfrac p 1, this correction is exact. ulated images are drizzled together, the rms noise in the final
In order to study the ability of Drizzle to remove the photo- photometry (which does not include any errors that could occur
metric effects of geometric distortion when pixfrac is not iden- because of missed or incorrectly identified cosmic rays) is
tically equal to 1, we created a 4-times–subsampled grid of ⱗ0.015 mag. Figure 4 displays the results of this process.
19 # 19 artificial stellar PSFs. This image was then blotted
onto four separate images, each with the original WF sampling 6. ASTROMETRY
but dithered in a four-point pattern of half-pixel shifts. As a We have also evaluated the effect of drizzling on astrometry.
result of the geometric distortion of the WF camera, the stellar The stellar images described in the previous section were again

2002 PASP, 114:144–152

150 FRUCHTER & HOOK

drizzled using the HDF shifts as above, setting scale p 0.5

and pixfrac p 0.6. Both uniform weight files and cosmic-ray
masks were used. The positions of the drizzled stellar images
were then determined with the “imexam” task of IRAF, which
locates the centroid using the marginal statistics of a box about
the star. A box with side equal to 6 output pixels, or slightly
larger than twice the FWHM of the stellar images, was used.
An rms scatter of the stellar positions of ∼0.018 input pixels
about the true position was found for the images created with
uniform weight files and the cosmic-ray masks. However, we
find an identical scatter when we downsample the original 4-
times–oversampled images to the 2-times–oversampled scale
of the test images. Thus, it appears that no additional meas-
urable astrometric error has been introduced by Drizzle. Rather,
we are simply observing the limitations of our ability to cen-
troid on images that contain power that is not fully Nyquist
sampled even when using pixels one-half the original size.

7. NOISE IN DRIZZLED IMAGES

7.1. The Nature of the Problem
Drizzle frequently divides the power from a given input pixel Fig. 5.—Schematic view of the distribution of noise from a single input
between several output pixels. As a result, the noise in adjacent between neighboring output pixels. See text for a more complete description.
pixels will be correlated. Understanding this effect in a quan-
titative manner is essential for estimating the statistical errors
7.2. The Calculation
when drizzled images are analyzed using object detection and
measurement programs such as SExtractor (Bertin & Arnouts In general, the correlation between pixels, and thus R , de-
1996) and DAOPHOT (Stetson 1987). pends on the choice of drizzle parameters and geometry and
The correlation of adjacent pixels implies that a measurement orientation of the dither pattern and often varies across an
of the noise in a drizzled image on the output pixel scale un- image. While it is always possible to estimate R for a given
derestimates the noise on larger scales. In particular, if one block set of drizzle parameters and dithers, in the case where all
sums a drizzled image by N # N pixels, even using a proper output pixels receive equivalent inputs (in both dither pattern
weighted sum of the pixels, the per-pixel noise in the block and noise, although not necessarily from the same input im-
summed image will generally be more than a factor of N greater ages), the situation becomes far more analytically tractable. In
than the per-pixel noise of the original image. The factor by this case, calculating the noise properties of a single pixel gives
which the ratio of these noise values differs from N in the limit one the noise properties of the entire image.
as N ⇒ ⬁ we refer to as the noise correlation ratio, R. Consider then the situation when pixfrac, p, is set to zero.
One can easily see how this situation arises by examining There is then no correlated noise in the output image—since
Figure 5. In this figure we show an input pixel (broken up into a given input pixel contributes only to the output pixel that lies
two regions, a and b) being drizzled onto an output pixel plane. under its center, and the noise in the individual input pixels is
Let the noise in this pixel be e, and let the area of overlap of assumed to be independent. Let dxy represent a pixel from any
the drizzled pixel with the “primary” output pixel (shown with of the input images, and let C be the set of all dxy whose centers
a heavier border) be a and the areas of overlap with the other fall on a given output pixel of interest. Then it is simple to
three pixels be b1, b2, and b3, where b p b1 ⫹ b 2 ⫹ b 3 and show from equations (4) and (5) that the expected variance of
a ⫹ b p 1. Now the total noise power added to the image the noise in that output pixel, when p p 0, is simply
variance is, of course, e2; however, the noise that one would
measure by simply adding up the variance of the output image 冘d xy苸C wxy2 s 4jxy2
(冘d xy苸C wxy)2
jc2 p , (6)
pixel by pixel would be

(a 2 ⫹ b12 ⫹ b 22 ⫹ b 33 )e 2 ! e 2. where jxy is the standard deviation of the noise distribution of

the input pixel dxy. We term this jc because it is the standard
The inequality exists because all cross terms (ab1, ab2, b1b2, deviation calculated with the pixel values added only to the
…) are missed by summing the squares of the individual pixels. pixels on which they are centered.
These terms, which represent the correlated noise in a drizzled Now let us consider a drizzled output image where p 1 0.
image, can be significant. In this case, the set of pixels contributing to an output pixel

2002 PASP, 114:144–152

 DRIZZLE: RECONSTRUCTING UNDERSAMPLED IMAGES 151

will include not only pixels whose centers fall on the output Although R must be calculated for any given set of dithers,
pixel but also those for which a portion of the drop lands on there is perhaps one case that is particularly illustrative. When
the output pixel of interest even though the center does not. one has many dithers and these dithers are fairly uniformly
We refer to the set of all input pixels whose drops overlap with placed across the pixel, one can approximate the effect of the
a given output pixel as P and note that C O P. The variance dither pattern on the noise by assuming that the dither pattern
of the noise in a given output pixel is then is entirely uniform and continuously fills the output plane. In
this case, the above sums become integrals over the output
冘d xy苸P axy2 wxy2 s 4jxy2 pixels, and thus it is not hard (though somewhat tedious) to
(冘d xy苸P axywxy)2
jp2 p , (7) derive R. If one defines r p p/s, where p p pixfrac and
s p scale, then in the case of a filled uniform dither pattern
where axy is the fractional area overlap of the drop of input data one finds, if r ≥ 1,
pixel dxy with the output pixel o. Here we choose the symbol
jp to represent the standard deviation calculated from all pixels
that contribute to the output pixel when pixfrac p p . The degree
Rpr Z (1 ⫺ 3r1 ) , (9)

to which jp2 and jc2 differ depends on the dither pattern and
the values of p and s. However, as more input pixels are av- and if r ≤ 1,
eraged together to estimate the value of a given output pixel
in P than in C, jp2 ≤ jc2. When p p 0, jp is by definition equal
to jc.
Rp1 Z (1 ⫺ 3r ) . (10)

Now consider the situation where we block average a region

of N # N pixels of the final drizzled image, doing a proper
weighted sum of the image pixels. This sum is equivalent to Using the relatively typical values of p p 0.6 and s p 0.5,
having drizzled onto an output image with a scale size Ns. one finds R p 1.662. This formula can also be used for block
However, as Ns k p, this approaches the sum over C, or, in summing the output image. For example, a weighted block sum
the limit of large N, Njc. However, a prediction of the noise of N # N pixels is equivalent to drizzling into a single pixel of
in this region, based solely on a measurement of the pixel-to- size Ns. Therefore, the correlated noise in the blocked image can
pixel noise, without taking into account the correlation between be estimated by replacing s with Ns in the above expressions.
pixels would produce Njp. Thus, we see that
8. SOME EXAMPLES OF THE APPLICATION
jc OF DRIZZLE
Rp .
jp Drizzle has now been widely used for many astronomical
image combination problems. In this section we briefly note
some of these and provide references where further information
One can therefore obtain R for a given set of drizzle pa- may be obtained.
rameters and dither pattern by calculating jc and jp and per- Drizzle was developed for use in the original HDF-N, a project
forming the division. However, there is a further simplification to image an otherwise unexceptional region of the sky to depths
that can be made. Because we have assumed that the inputs to far beyond those of previous astronomical images. Exposures
each pixel are statistically equivalent, it follows that the weights were taken in four filter bands from the near-ultraviolet to the
of the individual output pixels in the final drizzled image are near-infrared. The resulting images are available in the pub-
independent of the choice of p. To see this, notice that the total lished astronomical literature (Williams et al. 1996) as well as
weight of a final image (that is, the sum of the weights of all from the Space Telescope Science Institute via the World Wide
of the pixels in the final image) is independent of the choice Web.3
of p. Ignoring edge pixels, the number of pixels in the final Subsequently, Drizzle has also been applied to the Hubble
image with nonzero weight is also independent of the choice Deep Field–South (HDF-S; Williams et al. 2000). In this case
of p. Yet because the fraction of pixels within p of the edge it was used for the combination of images from NICMOS
scales as 1/N and the weight of an interior pixel cannot depend (A. S. Fruchter et al. 2002, in preparation) and STIS (Gardner
on N, we see that the weight of an interior pixel must also be et al. 2000) as well as WFPC2 (Casertano et al. 2000). In order
independent of p. As a result, 冘 d xy苸C wxy p 冘 d xy苸P axy2 wxy. to obtain dithered NICMOS and WFPC2 images in parallel
Therefore, we find that with STIS spectroscopy, HST was rotated, as well as shifted,
between observations during the HDF-S. All the software de-
jc2 冘dxy苸C wxyjxy 2 2
veloped to handle such challenging observations is now pub-
R2 p
jp2
p
冘dxy苸P axy2 wxy2 jxy2 . (8) licly available (see § 9).

3
See http://www.stsci.edu/ftp/science/hdf/hdf.html.

2002 PASP, 114:144–152

152 FRUCHTER & HOOK

The HDF imaging campaigns are atypical because they had ancillary tasks for assisting with determining the shifts between
a large number of dither positions. A more usual circumstance, images and the combination of WFPC2 data are available as
matching that described in § 3, is the processing of a small part of the “dither” package in STSDAS.
number of dithers without multiple exposures at the same point- We are continuing to improve Drizzle, to increase both ease
ing. A good example of such imaging and its subsequent pro- of use and generality. New versions of Drizzle will be incor-
cessing is provided in Fruchter et al. (1999), where gamma- porated into STSDAS software updates. Additional capabilities
ray burst host galaxies were observed using the STIS and will soon make the alignment of images simpler and will pro-
NICMOS HST cameras to obtain morphological and photo- vide the user with a choice of drizzling kernels, including ones
metric information. Similarly, Bally, O’Dell, & McCaughrean designed to speed up the image combination with minimal
(2000) have used Drizzle to combine dithered WFPC2 images change to the output image or weight—an enhancement that
with single exposures at each dither position in a program to may prove particularly useful in the processing of ACS images.
observe disks, microjets, and windblown bubbles in the Orion Although these additions may make Drizzle somewhat more
Nebula. flexible, the basic algorithm described here will remain largely
Examination of these published images may help the reader unchanged because it provides a powerful, general algorithm
to obtain a feeling for the results of using the Drizzle program. for the combination of dithered undersampled images.
In addition, an extensive set of worked examples of combining
dithered data using Drizzle is available in the Dither Handbook
(Koekemoer et al. 2001) distributed by STScI. Drizzle was developed originally to combine the HDF-N data
sets. We wish to thank our colleagues in the HDF-N team, and
9. CONCLUSION Bob Williams in particular, for encouraging us and for allowing
us to be part of this singularly exciting scientific endeavor. We
Drizzle provides a flexible, efficient means of combining also thank Ivo Busko for his work on the original implemen-
dithered data that preserves photometric and astrometric ac- tation of the STSDAS dither package, Bill Sparks for the ef-
curacy, obtains optimal signal-to-noise ratio, and approaches ficient algorithm used to calculate pixel overlaps, Hans-Martin
the best resolution that can be obtained through linear recon- Adorf for many entertaining and thought-provoking discussions
struction. An extensively tested and robust implementation is on the theory of image combination, and Stefano Casertano
freely available as an IRAF task as part of the STSDAS package for inciting us to develop a more general theory of the cor-
and can be retrieved from the Space Telescope Science Institute related noise in drizzled images. Finally, we are grateful to
World Wide Web site.4 In addition to Drizzle, a number of Anton Koekemoer for a careful reading of the text and to our
referee, Tod Lauer, for numerous suggestions, which signifi-
4
See http://www.stsci.edu. cantly improved the clarity and presentation of this paper.

REFERENCES

Bally, J., O’Dell, C. R., & McCaughrean, M. J. 2000, AJ, 119, 2919 Richardson, B. H. 1972, J. Opt. Soc. Am., 62, 55
Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393 Stetson, P. B. 1987, PASP, 99, 191
Casertano, S., et al. 2000, AJ, 120, 2747 Storrs, A., Hook, R., Stiavelli, M., Hanley, C., & Freudling, W. 1999,
Cowie, L. L., Hu, E. M., & Songaila, A. 1995, AJ, 110, 1576 NICMOS Instrument Science Report 99-005 (Baltimore: STScI)
Fruchter, A. S., et al. 1999, ApJ, 516, 683 Tody, D. 1993, in ASP Conf. Ser. 52, Astronomical Data Analysis
Gardner, J. P., et al. 2000, AJ, 119, 486 Software and Systems II, ed. R. J. Hanisch, R. J. V. Brissenden, &
Gull, S. F., & Daniell, G. J. 1978, Nature, 272, 686 J. Barnes (San Francisco: ASP), 173
Hanisch, R. J., & White, R. L. 1994, The Restoration of HST Images Trauger, J. T., et al. 1994, ApJ, 435, L3
and Spectra–II (Baltimore: STScI/NASA)
Weir, N., & Djorgovski, S. 1990, in The Restoration of HST Images
Koekemoer, A. M., et al. 2001, The HST Dither Handbook (Baltimore:
STScI) and Spectra, ed. R. L. White & R. J. Allen (Baltimore: STScI), 31
Lauer, T. R. 1999a, PASP, 111, 227 Whitmore, B., & Heyer, I. 1997, HST WFPC2 Instrument Science
———. 1999b, PASP, 111, 1434 Report 97-08 (Baltimore: STScI)
Lucy, L. B. 1974, AJ, 79, 745 Williams, R. E., et al. 1996, AJ, 112, 1335
Lucy, L. B., & Hook, R. N. 1992, in ASP Conf. Ser. 25, Astronomical ———. 2000, AJ, 120, 2735
Data Analysis Software and Systems, ed. D. M. Worrall, C. Bie-
mesderfer, & J. Barnes (San Francisco: ASP), 277

2002 PASP, 114:144–152