SuperNovae Classification from Telescope Images

Overview

The volume of data obtained from modern astronomical imaging techniques (e.g. nightly automated scans from large telescopes) is too large for direct human inspection. While some objects are stable and will be visible in many images, others are transient and only visible in some images. Supernovae are in the latter class, and reliable detection of them requires continual processing of nightly images. This project explores automation of the process, using some hand-categorized images as training data. From the training data, the goal is to train automatic classifiers that are as accurate as possible and reasonably fast.

Questions

The primary question is a binary classification "is this image from a supernova"? Using hand-labelled data you will train a classifier and test it using cross-validation.

The Dataset

The initial dataset is approximately 2GB and comes from Prof. Peter Nugent's group Links to an external site..

Reference

A recent paper on analysis of the dataset is available from this discussion link.

Tools

The current state-of-the-art for this problem uses special featurization and a random forest classifier. You can experiment further with random forests, although we recommend exploring some other approaches, e.g. either train a DNN using Caffe, or try condensed KNN.