Recently, researchers at the University of California, Santa Cruz, developed a deep learning model called Morpheus that allows pixel-level analysis of astronomical image data to identify and classify all galaxies and stars. The study was published May 12 in the Astrophysical Journal Supplement.
According to the paper, when working with images of an area of the sky, “Murphys” will generate a new image. In the figure, the algorithm model color-codes the celestial bodies according to their morphology, separates them from the background, and identifies stars and different types of galaxies.
In terms of accuracy, the paper notes that the false positive rate (false positive rate, FPR) of the “Murphys” model is only about 0.09 percent, compared to the catalog provided by the Hubble Space Telescope’s CANDELS project and the 3D-HST program. The false positive rate is the percentage of actual non-objects, but recognized as celestial bodies.
“Murphys” classified the celestial bodies and produced a color synthesis map.
Brant Robertson, a professor of astronomy and one of the two co-authors of the paper and head of the Computational Astrophysics Research Group at the University of California, Santa Cruz, said that with the rapid expansion of astronomical data sets, traditionally astronomer-completed tasks need to be automated.
“As human beings, there are things we simply can’t do,” he says. It is therefore necessary to find a way to use computers to process the vast amounts of data collected by large astronomical projects in the coming years”.
Take LSST, a large integrated sky telescope under construction in Chile, for example, and when completed, the project will take 800 panoramic images per night with a 3.2-megapixel camera, recording the sky twice a week. The vast amount of data generated by such large-scale astronomical projects far exceeds the analytical capabilities of astronomers.
In fact, there have been astronomers using deep learning to classify galaxies, but often using off-the-shelf image recognition algorithms that require additional data-collating work in advance. The difference between Morpheus is that it was created to process astronomical image data, and astronomers could enter raw image data in a standard digital file format.
“Murphys” uses the image segmentation and desynthetic process.
Another big advantage of “Murphys” is the pixel classification. When using other models, astronomers must first know what the image contains, and then provide the model with an image that then categorizes the entire galaxy, but “Murphys” can help astronomers discover galaxies and analyze them in pixels, according to Brent Robertson. This feature gives Morpheus the ability to process complex images.
The results of the classification of goods-South Field (Southern Sky region of the Great Observatory’s Cosmic Origin Depth Survey) of the Great Observatory.
Similar to the training process for other large deep learning models, the researchers “fed” a large amount of learning material to the model. The CANDELS observations, mainly provided by the Hubble Space Telescope, were analyzed and classified by dozens of professional astronomers and divided into about 10,000 galaxy clusters.
“Murphys” neural network architecture.
“Murphys” is an example of automatic morphological decomposition.
After studying, the researchers applied “Murphys” to the most complete and comprehensive cosmic map to date, the Hubble Heritage Field (HLF). The cosmic images are stitched together from 7,500 images of the stars taken by the Hubble Space Telescope over a 16-year period and contain about 265,000 galaxies.
To help deep learning models quickly complete a pixel-by-pixel analysis of the entire data set, the researchers configured the University of California, Santa Cruz’s supercomputer “lux.” The superplan was funded by the National Science Foundation for $1.547 million and covers astrophysics and climate science.