Better Input, Better Output: Improving photometric redshifts by enhancing training data and optimizing observations
Kalmbach, John Bryce
MetadataShow full item record
We are at the beginning of an era of large scale survey astronomy where we will soon measure photometry for billions of galaxies. In order to effectively use these galaxies for dark energy measurements we require measurements of the distances to these galaxies. Spectroscopic redshifts are not feasible for more than a small fraction of these galaxies and thus our primary distance measurements will rely on photometric redshift methods. This thesis highlights three challenges in photometric redshift estimation and techniques we developed to tackle these challenges: Using Information Theory to Optimize Bandpasses for Photometric Redshifts: We apply ideas from information theory to create a method for the design of optimal filters for photometric redshift estimation. We show the method applied to a series of simple example filters in order to motivate an intuition for how photometric redshift estimators respond to the properties of photometric passbands. We then design a realistic set of six filters covering optical wavelengths that optimize photometric redshifts for z <= 2. We create a simulated catalog for these optimal filters and use our filters with a photometric redshift estimation code to compare to the filters for the Large Synoptic Survey Telescope (LSST) which have key features in common with our optimal filters. Expanding Template Sets for Template Based Photo-Z Algorithms: Measuring the physical properties of galaxies such as redshift frequently requires the use of Spectral Energy Distributions (SEDs). SED template sets are, however, often small in number and cover limited portions of photometric color space. Here we present a new method to estimate SEDs as a function of color from a small training set of template SEDs. We first cover the mathematical background behind the technique before demonstrating our ability to reconstruct spectra based upon colors and then compare to other common interpolation and extrapolation methods. When the photometric filters and spectra overlap we show reduction of error in the estimated spectra of over 65% compared to the more commonly used techniques. We also show an expansion of the method to wavelengths beyond the range of the photometric filters. Finally, we demonstrate the usefulness of our technique by generating 50 additional SED templates from an original set of 10 and applying the new set to photometric redshift estimation. We are able to reduce the photometric redshifts standard deviation by at least 22.0% and the outlier rejected bias by over 86.2% compared to original set for z <= 3. Color Space Data Augmentation for Photometric Redshifts: When training sets for machine learning methods are not representative of the test set then there can be errors in the resulting estimates. In photometric redshifts this can happen when the color space of the spectroscopic data does not match the observed galaxy color space for an empirical photometric redshift estimation method. We first show how a lack of data in a region of color space of the training data affects photometric redshift estimation and then develop three different methods to add in synthetic training data to the missing area to mitigate the errors. Our best performing method lowers the photo-z bias by 51% and reduces the outlier fraction by 9.6% in the test data that lies in the missing area of color space compared to an unrepresentative training catalog.
- Physics