Star-Galaxy Separation via Gaussian Processes with Model Reduction

Modern cosmological surveys such as the Hyper Suprime-Cam (HSC) survey produce a huge volume of low-resolution images of both distant galaxies and dim stars in our own galaxy. Being able to automatically classify these images is a long-standing problem in astronomy and critical to a number of different scientific analyses. Recently, the challenge of "star-galaxy" classification has been approached with Deep Neural Networks (DNNs), which are good at learning complex nonlinear embeddings. However, DNNs are known to overconfidently extrapolate on unseen data and require a large volume of training images that accurately capture the data distribution to be considered reliable. Gaussian Processes (GPs), which infer posterior distributions over functions and naturally quantify uncertainty, haven't been a tool of choice for this task mainly because popular kernels exhibit limited expressivity on complex and high-dimensional data. In this paper, we present a novel approach to the star-galaxy separation problem that uses GPs and reap their benefits while solving many of the issues traditionally affecting them for classification of high-dimensional celestial image data. After an initial filtering of the raw data of star and galaxy image cutouts, we first reduce the dimensionality of the input images by using a Principal Components Analysis (PCA) before applying GPs using a simple Radial Basis Function (RBF) kernel on the reduced data. Using this method, we greatly improve the accuracy of the classification over a basic application of GPs while improving the computational efficiency and scalability of the method.