THE CHALLENGE OF VISUALISING ASTRONOMICAL DATA

Ray P. Norris, Australia Telescope National Facility, CSIRO Radiophysics Laboratory, Australia

Abstract

Current astronomical instruments now routinely produce such large volumes of data that it becomes difficult for an astronomer to obtain the information he wants from the data. While the machines get faster and cheaper, and are certainly able to handle the load, and the astronomer's brain is certainly capable of interpreting the data, the interface between machine and brain prevents the astronomer from assembling the data in a coherent fashion in his mind, and so prevents him from being able to extract all the useful information which resides in his data.

This bottleneck between human and machine poses problems both for those users who want to obtain an intuitive, qualitative, understanding of their data, and also for those users who want to obtain quantitative results from their data. Here I discuss ways, such as volume rendering of spectral-line data cubes, of representing the data which attempt to overcome this bottleneck. Finally, I discuss the results that users have obtained so far using implementation of these techniques at the Australia Telescope., and the future directions which we will explore.

In this paper, I concentrate on examples drawn from spectral-line radioastronomy. However, the arguments I make are applicable to several disciplines, most obviously to spectral data at other wavelengths.

1) Introduction

Current astronomical instruments produce enormous volumes of data, all of which is potentially useful. Our available computing resources are generally quite capable of dealing with this volume of information, and we know that the human brain is quite capable of absorbing this volume of information and of producing new scientific knowledge from it. However, there exists a bottleneck between man and machine which prevents the user from assembling this information in his mind in a coherent and integrated way, and so prevents the user from extracting all the information from his data. As a user, I deplore this because I want to make the most of my hard-won data. As one of those involved in running a national facility, I deplore this because I want to maximise the scientific return from the investment made in the telescope and its support.

The visualisation of data cubes has four distinct functions.

• Obtain an intuitive understanding of the data
• See features of the data which would otherwise remain unnoticed
• Obtain quantitative results from the data
• Communicate the results, both qualitative and quantitative, to other people

The challenge, then, is this. How can we break down this conceptual barrier between the user and the data, so that the user can get full scientific value from this hard-earned data?

Intrinsic to the nature of this subject is the difficulty of describing some of the three-dimensional examples in two dimensions. At the ADASS meeting this was overcome by projecting movies demonstrating some of the techniques described here, but it is difficult to reproduce these in this paper! This illustrates a further challenge from the fourth function of visualisation: if, as I confidently predict, these visualisation techniques become widely-used in astrophysics, how do we publish the results?

2) Spectral-line data cubes

2.1) Definition of the problem

Figure 1 shows the three main ways in which a user might want to interact with her spectral-line cube. First there are simple quantitative measurements that she wants to make which can in principle be done with existing packages of software (AIPS, GIPSY, etc.). However, many users find that current tools are inadequate even for this simple task. While some work is being done to improve the capabilities of current software, it is likely that these demands will be met only by the release of new software, such as AIPS++ , over the next few years.

A second need is the visualisation of the cube, with which this paper is mainly concerned. A third need is for making quantitative measurements on the volume visualisation, but as this depends on a union of techniques from the first two branches of this diagram, this must be postponed until a satisfactory solution is found for these first two.

The image of a spectral line source, when imaged with the Australia Telescope (AT), consists of up to 8192 separate images, each of up to 8192 x 8192 pixels, and each of which portrays the emission at a different wavelength. However, this volume of data (2000 Gbyte!) is too large for our current computing resources to handle adequately, so more typically a user will produce 256 images each of which is 256 x 256 pixels. These separate images can conveniently be regarded as being planes of a cube containing 256x256x256 three-dimensional pixels, or voxels (volume pixels).

This cube, with one byte assigned to each voxel, occupies just 67 Mbyte, and so easily fits in the memory of a high-end workstation, and is easily manipulated by the available compute power. However, no adequate way exists to transfer this data to a user's brain, and so expensive information is being discarded.

A traditional technique of displaying this data (in the AIPS task TVMOVIE, for example) consists of showing a two-dimensional (2-D) representation of each plane of the cube, in sequence, so that the user sees a movie of one plane after another. If this is done sufficiently rapidly, then this technique can go some way towards giving the user an intuitive impression of the data.

Typically, this data will be of a spectral line from a source in which different parts of the object are moving with different velocities (and hence different wavelengths, because of Doppler shift), so that the shape and position of the image will vary from plane to plane of the cube. If there is a velocity gradient in a cloud of gas, for example, then the position of the maximum emission in each plane will shift steadily as different planes are viewed, so that the movie gives the appearance of an object moving across the screen. However, this technique does not reveal subtle effects to the user, and in the case of a large number of planes there may be a considerable time lag between viewing different parts of the movie, so a user will not be able to relate frames at the start of the movie to frames at the end of the movie. At the AT we are exploring a number of alternative techniques which present the data cube as a single three-dimensional (3-D) object.

2.2) Volume Rendering

The commonest form of 3-D visualisation seen in other disciplines is some form of surface shading. In this technique, objects are assumed to be opaque, and a colour, brightness, and reflectivity is assigned to each pixel along the surface of each object. Rays of light are then traced from some source of light to the users eye, resulting in a realistic life-like image if the objects are familiar ones taken from everyday life.

This technique is not well-suited to astronomical applications for two main reasons:

• We need to see the data inside an object (e.g. a galaxy) as well as on the surface.
• Astronomical data generally have a relatively poor signal-noise ratio so that the surfaces are not well-defined.

Despite these two disadvantages, surface rendering can be applied usefully to astronomical objects, such as the neutral hydrogen in a galaxy, by defining a surface of equal brightness. The level of this surface must be chosen well above the noise, and some smoothing may be necessary to reduce noise effects even further. Conventional ray-tracing techniques then portray the data as a solid object.

Volume rendering techniques are also employed in other disciplines. In volume rendering techniques, individual voxels of the cube are assigned a brightness, colour, and opacity, and then some form of ray-tracing is done through the volume of the cube. These techniques are well-known in medical and geophysical imaging , but the data in these disciplines differs from astrophysical data, and so the techniques differ too.

The algorithm which appears most successful in representing an astronomical data cube to users is the "hot gas" algorithm, which was first implemented at ATNF by Eric Greisen. In this algorithm, a look-up table is constructed which assigns colour, brightness, and opacity to each voxel, depending on the value of the data in that voxel. An approximate radiative transfer equation is then solved through the cube for each ray of light which will strike the users eye. The resulting image appears as a cube of glowing gas, and a galaxy will appear as a cloud of swirling mist within the cube.

2.3) An example: Visualisation of an OH/IR star

An OH/IR star is a red giant star which is shedding enormous amounts of material as a cool stellar wind. At one particular radius, maser emission occurs from the OH molecules in the ejected material, so that there is a thin spherical shell of OH maser emission surrounding the star. Because of the details of the maser process, strong emission is seen from the Earth only near the front and back of this shell, and barely any emission is seen from the limbs. The material is blown of isotropically, so that the radial velocity (as seen from the star) is the same at all points of the sphere.

The result is shown in Figure 2. The spectrum is a characteristic double peaked shape, with the strongest emission, which occurs at the maximum and minimum velocity with respect to the Earth, coming from the front and back of the shell. An image at these maximum and minimum velocities shows a single compact source in the centre of the image, whilst at intermediate velocities the emission comes from a ring. The geometry is such that line-of-sight velocity is exactly proportional to distance along the line of sight. Thus if we visualise the data cube, we see a three-dimensional sphere of emission (actually only parts of a sphere, because the emission from the limbs is very weak) sitting in our cube of data. This is one of the very rare occasions in astronomy when we can see an object in its true 3-D reality; generally astronomical objects are seen only as 2-D representations.

2.4) The advantages of 3-D representation

Conventional movie techniques (such as AIPS TVMOVIE) take too long for the eye/brain to associate information in different parts of the spectrum, and so do not allow the brain to get a intuitive impression of the data as a whole. Volume rendering techniques, on the other hand, show all the data at once, thus overcoming these problems, and so allow the brain to make full use of its powerful pattern-recognition capabilities. For example, a feature may be present in each plane of the cube but is indiscernible in the individual images because it is barely above the noise level, but when seen as a cube the feature may show up easily to the eye as a bar running through the cube. Similarly, subtle features such as a warp, or non-circular motions, in the neutral hydrogen emission from a galaxy will be far more obvious in a volume-rendered cube than in a TVMOVIE. An unexpected benefit of the volume visualisation is that imaging artefacts are also more easily seen (and thus corrected) on a volume-rendered image than in a TVMOVIE.

2.5) The use of motion and pseudo-interactive visualisation

Experiments with users at the ATNF have shown that viewing a rendered cube from one direction does not adequately convey the data to the users mind, but that motion is an important factor in enabling this data transfer. It appears that motion is an essential cue to the human brain in interpreting an image as a 3-D object. An object such as the OH/IR star described above is difficult to interpret if any one frame is shown statically, but is immediately visualised as a 3-D object if it is rotating Conversely, if the motion of a rotating 3-D cube is stopped, the subjective 3-D effect quickly disappears, and the regular structure apparent in the 3-D object is replaced by the 2-D appearance of an ensemble of patches of colour. It appears that the brain's powerful pattern recognition capabilities are best adapted to moving objects.

We therefore provide tools at the ATNF for rotating the image of the cube. However, the volume rendering is too slow (taking 5 s per frame for a 128x128x128 cube on a 400-Mflop machine) for large cubes, and so most users make a movie which may take an hour to compute, but can then be displayed on the workstation screen to give a realistic impression of the data. Such a movie then shows the cube rotating smoothly in space, with the object of interest appearing as a region of glowing gas sitting within the cube. Such movies appear to be powerful tools. However, it is likely that an even better intuitive impression of the data will be gained if the user could interactively rotate the cube. For example, she could use a spaceball to turn and twist the object, peering at it from all angles. Unfortunately, the compute power required for this is beyond the available resources.

An alternative to full interactive visualisation is a technique that I call pseudo-interactive visualisation, in which a movie is pre-calculated not of just the frames needed for rotation through one turn, but of every possible view of the cube from any angle over the sphere. If images are calculated at 2º intervals over the sphere, then the resulting 160 Mbyte can be stored in memory, and the images roamed using a spaceball or mouse, giving the user the impression of fully interactive motion of the cube. We are in the process of writing the software to do this.

2.6) Obtaining quantitative data from 3-D visualisation

When measuring the integrated flux coming from a region in a 2-D image, the region is easily specified using a mouse to "click-and-drag". In 3-D space, the problem becomes substantial.

One technique is to display three 2-D cross-sections of the cube, and to represent the mouse position in these. however, this method is neither intuitive nor easily tied in to the 3-D image with which the user will have gained his intuitive understanding of the data. Instead, we need to devise a way of moving a 3-D cursor within a 3-D cube.

Although it is straightforward to replace the mouse by a 3-D device such as a spaceball, and to replace the 2-D cursor by a 3-D cursor, the problem of locating the 3-D cursor inside a volume is a difficult one. If the cube is stationary, it is difficult for the user to judge depth within the cube. This can be solved by continuously rotating the cube, as in the movies discussed above, but it then becomes difficult for the user to coordinate his hand movements with the position of a cursor in a rotating cube. We have not yet solved this problem, and it is likely to be a problem that will require some ingenuity in its solution. When we have solved this problem of manipulating a 3-D cursor, we will then have to tackle the problem of selecting a 3-D region.

An additional technique that should enhance the 3-D effect is the use of stereoscopic vision. We are starting to implement this at the ATNF using specialised hardware.

3) Practical Experience: the ATNF visualisation project

3.1) Overview and Strategy

The ATNF visualisation project is a research project on which three people (R. P. Norris, T. Oosterloo, R. E. Gooch) are working, and which aims to explore the use of visualisation on astronomical data cubes. Because 3-D visualisation is compute-intensive, it currently requires specialised machines (typically hundreds of Mflop/s are required), and the required performance is obtained from these machines only by writing specialised code. However, the continuing increase in the power of affordable computers means that within a few years the techniques which are being pioneered now can be transferred to standard desktop workstations. At that stage we intend to port our techniques to applications in AIPS++ and other imaging packages, but for the present we have abandoned portability and use specialised prototype code on specialised machines (such as the 400-Mflop MVX visualisation accelerator using 5 i860 processors in parallel).

In addition to astronomical visualisation, the group has strong links into medical and geophysical imaging projects at the CSIRO Division of Radiophysics, so that a significant degree of cross-fertilisation takes place.

Aspects of astronomical visualisation currently being investigated at the ATNF include:

• Slicing to obtain a two-dimensional cross-section of a 3-D object, including slicing along curved surfaces
• Volume rendering, including the hot gas algorithm discussed above, surface rendering, and others.
• Use of stereoscopic devices for volume rendering
• Use of devices other than a mouse and trackball to interact with the data, such as a spaceball and dials.
• The problem of how to construct a 3-D cursor, and how to select a 3-D subsection of a moving 3-D volume.

In the future, we hope to extend this investigation to virtual reality techniques such as use of a powerglove with tactile feedback. For example, a user could try to push down a feature in his data if she thought that it was an imaging artefact. If it was poorly constrained by the data (e.g. corresponding to an information gap in the Fourier distribution of a synthesised image), it would be soft and easily pushed down. If, on the other hand, it was required to be consistent with the input data, then it would feel hard and difficult to push away.

This last example is also an example of the more general technique of hypothesis testing. There are cases both in astronomy and in geophysics where the end result is constrained relatively poorly by the data, so that a researcher typically wants to ask the question "How do I know whether to believe this particular feature of the data?", or, more generally, "How well is this feature constrained by the input data?". We hope to explore other techniques to allow the user to interact with the data in this way.

3.2) Results with Real Live Users

So far, we have offered ATNF users basic volume rendering facilities, typically using the hot gas algorithm, and invited them to make movies as described above. As always with new techniques, there is some investment of time and effort that needs to be made by the user before he can use these techniques successfully. This, coupled with software and documentation which isn't always as user-friendly as we would like, together with the natural conservatism of many users, means that only perhaps 50% of spectral-line users of the ATNF have chosen to use the visualisation facilities. Nevertheless, the results have been extremely encouraging. We have found that most users see more astrophysically significant features in their data after viewing it on the visualisation facilities than they did when viewing it with AIPS TVMOVIE or other tools.

Typically, a user will first examine her data using AIPS, decide on some interpretation of the data, and then, at a much later stage, will visualise it using our facilities. Users have then discovered such features as non-circular motions in galaxies, absorption holes in the centres of galaxies, and artefacts in the data which show up as lines in the cube, even though their significance in any one channel may be low. For example, in preparing the movie of an OH/IR star for this conference, I discovered a feature in the data that I had not seen before even though I was, I thought, very familiar with the data: a blob of emission on the nearside of the sphere which is clearly, in the movie, at a greater radius from the star than the rest of the emission. This blob had not been noticed in previous analyses of the data, and may indicate amplification of the stellar thermal emission by the OH masers along our line of sight to the star.

Despite this anecdotal evidence of the value of these visualisation techniques, we do not yet have any carefully controlled experiments to measure their effectiveness. It is likely that such experiments belong more properly to the realms of psychology or sociology than to astrophysics or computer science, and yet they are essential to the most effective development of the subject. Future work will have to break down these inter-disciplinary barriers if we are to find out which techniques are most likely to be effective.

3.3) The significance of the rendered phase space

In the special case of the OH/IR star discussed above, the cube seen by the user corresponds to real 3-D space, and there is no confusion between the image and reality. However, this is not generally the case, and in general a user views an object in a three-dimensional phase space consisting of two spatial axes (Right Ascension and Declination) and one velocity axis. The user must be careful, then , not to confuse the image he is seeing with the physical appearance of the object in space. For example, the image of neutral hydrogen from a galaxy appears to look like a real object in the rendered cube which could be interpreted mistakenly as the actual appearance of the galaxy.

Does this matter? It appears that many users are already accustomed to viewing phase-space representations such as longitude-velocity diagrams, and to these users the extension to 3-D is not worrisome. To these users, all the interpretation and modelling may be done in the x-y-v phase space, and only after the object is understood is some x-y-z representation deduced. To others, however, this possible confusion is troubling, and this may inhibit the usefulness of these techniques for these users.

4) Conclusion

I have shown here some of the problems and the successes in visualising 3-D data. The subject is probably in its infancy, and future work will encounter challenges not currently envisaged. We are currently limited by the available computing hardware, and it is likely that this will change over the next few years. Even virtual reality techniques are likely to become much more widespread as the technology becomes more common and affordable, largely due to the use of these techniques for home entertainment. Our challenge then will be to implement these techniques in such a way that astronomers find them easy and intuitive to use, and where possible make them available within standard packages (e.g. AIPS++) so that they will run on standard equipment that sits on the desk of the average astronomer.