MegaMorph - Short version
The MegaMorph (Measuring Galaxy Morphology) project has developed novel approaches to aid the decomposition of galaxy images into their constituent physical components.
MegaMorph is based on Galfit and Galapagos, two well established pieces of software. Both codes, however, have been changed in order to use multi-wavelength data simultaneously, enabling the fitting of fainter galaxies in a physically meaningful way. The effect of this are two-fold:
1) It allows to reliably fit galaxy light profiles both more accurately, more physically meaningful and for more galaxies than other current software, due to less catastrophic failures due to low signal-to-noise imaging. We have developed and tested the software mainly on GAMA data, but have successfully run it on multiple other datasets. It seems to work great and it will help to understand galaxy evolution better than anyone previously had.
2) As colour information is 'used' by the software, it allows more accurate separation of different galaxy components (e.g. generally blue disks from generally red bulges).
We provide public codes for Bulge-Disk-Decomposition of large numbers of galaxies. Many science questions are waiting to be answered.
The Team& Funding
The MegaMorph team mainly consists of myself, Steven Bamford (both Nottingham at the time), Marina Vika and Alex Rojas (both at Carnegie Mellon University in Qatar then). Over time, we have gathered valuable ‘associates’ that had big influence on our work, e.g. Marco Barden (the author of original Galapagos), Lee Kelvin (the author of SIGMA, a similar pipeline used in GAMA), Chien Peng (the author of the original Galfit), Benedetta Vulcani, Rebecca Kennedy and others. (Most of these on a cheesy photo on the right)
MegaMorph also has strong links with the GAMA project.
The project was primarily funded by a grant from the Qatar National Research Foundation, which provided two postdocs dedicated to tackling the problem, one at Nottingham and another at CMU-Q. Steven Bamford is supported by an STFC Advanced Fellowship. We have also received some funding from Amazon Web Services.
Image from DSS
The Problem & Idea
It is becoming clear that the striking difference between ellipticals and spiral galaxies is actually a result of variation in the relative prominence of their more fundamental spheroid and disk components. Our understanding of galaxies would therefore be greatly improved by considering these physical components separately. However, measuring the properties of the individual components within a galaxy is considerably more difficult than measuring its overall properties as done by profile fitting codes in their previous versions.
Previous fitting routines only used a small fraction of the available data (e.g. one band of a multi-band survey). Several independent fits on each band do not overcome this problem, simultaneous usage of all band and multi-component fitting is required. More, current 1-band-1-component fitting result are dependent on the chosen wavelength for the fit (as can be imagined from the upper figure on the left) due to the mixing of the light of 2 separate galaxy components with different colours and which dominate different regions of the galaxy image, e.g. a red bulge in the center and a blue disk in the outskirts (clearly visible in the lower image). The choice of different bands leads to different measured properties, and, ultimately, different conclusions.
Reliable Bulge-Disk-decomposition on one-band data, as carried out by several groups, is challenging because of parameter degeneracies and multiple minima in likelihood space. Additional colour information could eliminate some of these minima and make the fit more stable and physically meaningful, e.g. by effectively allowing the code to fit a red stellar population in the red bands and extrapolating to blue bands and vice versa (although this is not what our code actually does in practice).
The MegaMorph project was tackling this problem by utilizing the full set of multi-colour information available for each galaxy and so potentially is able to separate different stellar populations within a galaxy and derive physically-meaningful structural parameters.
The starting point
We use both Galapagos (Galaxy Analysis Large Areas: Parameter Assessment by Galfiting Objects from SExtractor; Barden et al, 2012) and Galfit3 (Peng et al., 2010), two pieces of established and well tested software, which we adapted to perform robust, physically meaningful galaxy bulge-disk decompositions using data from many wavelength bands simultaneously, while retaining backwards compatibility wherever possible.
Galapagos (Barden et al. 2011), is a wrapping script which, after an initial setup by the user, runs the entire fitting process without further user interaction:
Galapagos applies SExtractor (Bertin & Arnouts 1996) for object detection.
Using an intelligent method, Galapagos automatically decides which neighbouring galaxies have to be fit simultaneously and which neighbours can be masked out. It automatically creates mask images used by Galfit in the fitting process.
Dealing with galaxies in order of decreasing brightness, Galapagos:
writes out a Galfit start file & runs Galfit
reads in the fitting result and uses this for deblending purposes in the further process of the code
Both Galapagos and Galfit have been thoroughly tested by several independent groups (including myself). We carried out extensive tests of Galapagos ourselves, using both real and simulated data. We find that Galapagos in general returns very good results except for a small systematic offset that can be seen for the faintest galaxies with very high Sérsic indices that are most sensitive to uncertainties in the sky estimation. We were able to show the independence of galaxy parameters from both distance and magnitude of neighbouring objects as measured in simulated data. Whereas other fitting methods are sensitive to neighbours, Galapagos and Galfit are not.
Carnegie-Mellon-University in Doha, Qatar
What we did - Single Sérsic fits
The MegaMorph project has added many additional features to Galfit and Galapagos. We have:
adapted both Galfit and Galapagos to be able to use multi-wavelength data simultaneously.
tested these new versions on both simulated data (following Haeussler et al., 2007), real survey data (GAMA, Driver et al., 2011), and artificially redshifted real galaxies (using Ferengi).
sped up the code in critical places and nearly halved the CPU time needed to run on real data.
have importantly enabled Galapagos to use variable PSFs, depending on objects position. This is important for ground-based and large-scale surveys.
added a selected target list to the code in order to save fitting time if a user is only interested in a subset of the objects in the FOV.
We have successfully run the codes on one region of the GAMA survey with several 10s of thousands of galaxies observed in 9 bands. 3 of the fits are shown on the left. In this run, only the magnitude varies (with complete freedom) from band to band, the other parameters are constrained to some sensible/physical polynomial. As one can see, the fitting magnitudes (red triangels in the left column) are in better agreement with photometric data (blue squares and stars) than single-band fits (yellow crosses), and the galaxy sizes recovered in each band provide a much smoother and more sensible variation than single-band values (white line vs yellow crosses in the right column). The degree of the polynomial used is user specified, the code offers full flexibility on each parameter individually. In fact, we usually run the fitting process with size and Sérsic index being polynomials of second order as a function of wavelength. These results have been published in Häußler et al. (2013, MNRAS, 430, 330) and have been used in Vulcani et al. (2014, MNRAS, 441, 1340) and Kennedy et al. (2015, MNRAS, 454, 806). The multi-wavelength Galfit version will be presented in Bamford et al. (in prep).
As a different approach (Vika et al., 2012, MNRAS, 435, 623), we have successfully redshifted a sample of ~160 local galaxies to redshifts out to z=0.25 and have fit all these images with both the original and the multi-wavelength version of Galfit in order to understand observational biases introduced by reduced image resolution and cosmological dimming at higher redshifts on measured galay properties in real data.
What we did - 2-component fits
We have further introduced a second fitting component, effectively moving from single Sérsic profile fitting to full Bulge-Disk-Decomposition. We find that multi-wavelength data and fitting is able to overcome many of the degeneracy problems that current, 1-band, B/D compositions have, due to its power of using the full colour information of a galaxy image on a pixel to pixel basis. In the future, an accurate best-model-selection has to be implemented and employed to automatically choose which fit resembles the galaxies profile more accurately. This is work in progress.
We are currently writing up the results from Bulge-Disk fitting and can show that the multi-band codes can separate two different galaxy components better than single-band approaches (Haeussler, 2016, in prep). The results of the tests will be published in an upcoming paper, but have already been used in Kennedy et al. (submitted). On the artificially redshifted sample, we have already shown that the multi-band fitting significantly improves accuracy and stability of Bulge-Disk decompositions (Vika et al., 2014, MNRAS, 444, 3603).
To control the computational intensity of the task, we have tried to use efficient algorithms and tools to optimally use the full CPU available. We are still trying to speed up the actual code in places, and have adapted it to work on high-performance computer facilities (at least in parts) either on local HPC machines available to the user or e.g. Amazon Web Services.
Testing has been done and demonstration papers have been or will soon be published. We have published the codes and we invite everyone to use them on their dataset with their own setups.
Bonus: Non-Parametric Components and MCMC
Galaxies are complex structures and, beyond the general distinction between spheroids and disks, they display a range of higher level features that make it difficult for computational methods to extract meaningful information. To overcome this problem, we have introduced non-parametric components into the fit to account for the rich variety of galaxy features which ‘distract’ conventional model-fitting methods.
Finally, we have made it possible to fully sample the parameters’ posterior probability space with the aims of (a) assuring the robustness of the approach, (b) quantifying parameter confidence intervals and degeneracies, and (c) performing reliable model selection. For this, we have used a MCMC/MultiNest approach.
Super Bonus: Running on IFU data
Finally, we have recently developed a way to run this method on IFU data (Johnston et al, in prep). One can see IFU data as a spectrum at each position, or as a series of many many images of the same objects at different wavelengths, the latter one being exactly what GalfitM needs. Due to CPU time constraints, it is impossible to run GalfitM on the IFU data directly, it would simply take too long. Evelyn Johnston has developed a way to make this idea work nontheless. The method works very well and will be published soon (code, too, once it has a name) for everyone to use on their own data.
From tests and comparison with other analyses, this method works very well and enables the separation of the spectra of a galaxy's disk and bulge, making an independent analysis of their chemical abundances, stellar population ages, metallicities, etc. feasible. The code still takes quite a long time per galaxy to cover the full wavelength range, so it would need another wrapper script to fully automate it on large samples of galaxies, but on an individual bases, the separation works very well. The top plot on the right shows the brightness of a galaxy in black, the decomposed bulge and disc spectra in red and blue respectively, and purple is the combination of the 2. From these spectra, it is possible to derive ages and metallicities of the bulge and disk (lower left plot, bulge in red, disk in blue). The right plot shows the radial trend as derived on the smooth and noiseless bulge and disk models. A comparison with a pixel-by-pixel analysis (lower middle plot) shows the same radial trend, but with much higher noise levels.