Spiral, barred, elliptical and irregular: Computers automatically classify galaxy shapes

Jun 02, 2010 by Robert Massey
A picture of the Abell Cluster taken using the Hubble Space Telescope. The picture encapsulates the diversity in galaxy types observed in our Universe. We can see a giant elliptical galaxy at the centre of the cluster, a beautiful spiral in the bottom right-hand corner and many smaller systems displaying a wide range of shapes, sizes and colours. Credit: NASA/ESA and the Hubble Heritage Team (STScI/AURA).

(PhysOrg.com) -- Scientists at University College London and the University of Cambridge have developed machine-learning codes modelled on the human brain that can be used to classify galaxies accurately and efficiently. Remarkably, the new method is so reliable that it agrees with human classifications more than 90% of the time. The research will appear in a paper in the journal Monthly Notices of the Royal Astronomical Society.

There are billions of in the Universe, containing anything between ten million and a trillion stars. They display a wide range of shapes, from elliptical and spiral to much more irregular systems. Large observational projects - such as the Sloan - are mapping and imaging a vast number of galaxies. As part of the process of using these data to better understand their origin and evolution, the first step is to classify the types of galaxies within these large samples. The 250,000 members of the public participating in the Galaxy Zoo project recently classified 60 million such galaxies by eye.

Now, a team of astronomers has used Galaxy Zoo classifications to train a known as an artificial neural network to recognize the different galaxy types. The artificial neural network is designed to simulate a biological neural network like those found in living things. It derives complex relationships between inputs such as the shapes, sizes and colours of astrophysical objects and outputs such as their type, mimicking the analysis carried out by the human brain. This method managed to reproduce over 90% of the human classifications of galaxies.

“We were astonished that a computer could do so well” says Dr Manda Banerji from the Institute of Astronomy at the University of Cambridge who led the research, which formed part of her PhD thesis at UCL. “This kind of analysis is essential as we are now entering a new age of astronomical surveys. Next generation telescopes now under construction will image hundreds of millions and even billions of galaxies over the coming decade. The numbers are overwhelming and every image cannot viably be studied by the human eye.”

A large-scale sky survey in which the UK is playing a leading role is the Dark Energy Survey (DES) due to commence in 2011, which is expected to image 300 million galaxies over 5 years. Another survey called the VISTA Hemisphere Survey being led by astronomers at the University of Cambridge, has just started taking data and will image galaxies over the entire southern hemisphere.

Professor Ofer Lahav, head of Astrophysics at UCL and chair of the international DES Science Committee, who supervised Banerji’s thesis, commented: “While human eyes are very efficient in recognizing patterns, clever computational techniques that can reproduce this behaviour are essential as we begin to push the boundaries of our observable Universe and detect more distant galaxies. This study is an important step in that direction.”

Explore further: How baryon acoustic oscillation reveals the expansion of the universe

More information: The work appears in “Galaxy Zoo: Reproducing Galaxy Morphologies Via Machine Learning”; Banerji M., Lahav O., Lintott C. J., Abdalla F. B., Schawinski K., Bamford S. P., Andreescu D., Murray P., Raddick M. J., Slozar A., Szalay A., Thomas D. and Vandenberg J., Monthly Notices of the Royal Astronomical Society, in press. A pre-print of the paper can be found at arxiv.org/abs/0908.2033 .

Related Stories

Antennae Galaxies

May 19, 2008

This image of the Antennae galaxies is the sharpest yet of this merging pair of galaxies. During the course of the collision, billions of stars will be formed. The brightest and most compact of these star ...

Arecibo Begins Search for Dark Galaxies

Feb 08, 2005

Fitted with a new compound eye, the Arecibo radio telescope in Puerto Rico last week began a multiyear effort to survey all the galaxies in a large swath of sky out to a distance of 800 million light years—a ...

Colliding galaxies make love, not war

Oct 17, 2006

A new Hubble image of the Antennae galaxies is the sharpest yet of this merging pair of galaxies. As the two galaxies smash together, billions of stars are born, mostly in groups and clusters of stars. The ...

Recommended for you

The Great Cold Spot in the cosmic microwave background

21 hours ago

The cosmic microwave background (CMB) is the thermal afterglow of the primordial fireball we call the big bang. One of the striking features of the CMB is how remarkably uniform it is. Still, there are some ...

Mystery of rare five-hour space explosion explained

Sep 17, 2014

Next week in St. Petersburg, Russia, scientists on an international team that includes Penn State University astronomers will present a paper that provides a simple explanation for mysterious ultra-long gamma-ray ...

User comments : 4

Adjust slider to filter visible comments by rank

Display comments: newest first

PinkElephant
3 / 5 (1) Jun 02, 2010
But how much work does this really save, if to catch the 10% incorrect identifications, a human reviewer would still have to look over every single classification?

Which begs another question: what's the point?
Adriab
not rated yet Jun 03, 2010
generally classifiers give you an a percent-likelihood that a given sample is in each class, so the lower the mean difference in percent-likelihood of the top ranking class to the lower ranking ones will relate to a higher percent-error.
These algorithms are all about the computer being able to grind through a large percent of the data, you take the classifications it gives when it has a high certainty, and manually check when it gives a low certainty.
For a well-made classifier and good data, the low certainty classifications are not too common. Then again, image processing/classifying is a hell of a tricky task.
yyz
not rated yet Jun 03, 2010
Actually, this was one of the goals of the Galaxy Zoo project...to use data from galaxy classification (by humans) to improve galaxy classification (by computer). As noted in the article, the floodgates are about to open in terms of raw data with the launch of large scale survey programs (DES, VISTA Hemisphere Survey, PanSTARRS and LSST, to name a few). Teaching machines to better sift through this data (galaxy classification being only one example) can only help scientists trying to build data sets that are both accurate and reliable.
Birger
not rated yet Jun 03, 2010
Artificial neural networks are also potentially useful in interpreting X-ray images when searching for breast cancer -I hope this algorithm can be put to this use as well.
A problem with artificial neural networks is that they cannot be saled up -a problem our brains have solved by evolution. Unfortunately the article does not mention if they have made progress with the scaling problem.