Harnessing the power of CNNs for unevenly-sampled light-curves using Markov Transition Field

Abstract

The search for exoplanets has evolved from case by case data inspection to automatic pattern recognition methods for processing a very large number of light curves. For this reason, the use of machine learning techniques has become a common practice in the field, where deep learning models are now in the spotlight as a promising leap forward towards automation. However, despite being faster than manual inspection, they usually still need hand-crafted features to achieve good results. Moreover, not all methods allow real world data where a large portion of the data is missing or at least is not regularly sampled. In this paper, we propose a method that only requires the raw light curve to identify exoplanets without the need of additional metadata or specific formats for the time series. We transform unevenly-sampled time series (light curves) of variable length into a 2-channel fixed-sized image using Markov Transition Field, which feeds a convolutional neural network that classifies candidate transients. We conducted experiments using the Kepler Mission dataset, identifying two key results: (1) the method is competitive in terms of performance to the state-of-the-art alternatives, yet it is simpler and faster. (2) A Markov Transition Field can be used as an effective stand-alone data product for analyzing unevenly-sampled transient light curves.