How to Find More Supernovae with Less Work: Object Classification Techniques for Difference Imaging

TL;DR

This paper demonstrates that advanced machine learning techniques like Boosted Decision Trees, Random Forests, and Support Vector Machines significantly improve the discrimination of supernovae from artifacts in difference imaging, reducing false positives and increasing detection efficiency.

Contribution

The paper introduces the application of sophisticated machine learning classifiers to difference imaging for supernova detection, outperforming simple threshold methods.

Findings

Reduced non-supernova candidates by a factor of 10.

Increased supernova identification efficiency.

Enhanced automated transient detection pipelines.

Abstract

We present the results of applying new object classification techniques to difference images in the context of the Nearby Supernova Factory supernova search. Most current supernova searches subtract reference images from new images, identify objects in these difference images, and apply simple threshold cuts on parameters such as statistical significance, shape, and motion to reject objects such as cosmic rays, asteroids, and subtraction artifacts. Although most static objects subtract cleanly, even a very low false positive detection rate can lead to hundreds of non-supernova candidates which must be vetted by human inspection before triggering additional followup. In comparison to simple threshold cuts, more sophisticated methods such as Boosted Decision Trees, Random Forests, and Support Vector Machines provide dramatically better object discrimination. At the Nearby Supernova Factory, we reduced the number of non-supernova candidates by a factor of 10 while increasing our supernova identification efficiency. Methods such as these will be crucial for maintaining a reasonable false positive rate in the automated transient alert pipelines of upcoming projects such as PanSTARRS and LSST.