Authors: Leah Zuckerman (University of Colorado, Boulder / NSO), Kevin Reardan (University of Colorado, Boulder / NSO)

To answer many open questions in solar research, we need to identify and segment important features in solar imaging data. Current techniques for feature identification rely on ad-hoc codes or by hand analysis, and these can be slow, computationally inefficient, and unreliable. For example, the primary features of interest within the solar photosphere are granules, inter-granular lanes, and facular bright points (small areas of high flux between granules, caused by strong magnetic field). However, there is currently no absolute definition of a where a granule ends and an intergranular lane begins, or how small and bright a region must be to be considered a facular bright point. Instead, experts define these regions with simple, empirically-driven, rule-based algorithms. While these heuristic methods are useful in some contexts (counting granules, tracking granule lifetimes, etc.), they can provide wildly inconsistent results, and are highly sensitive to variability in image resolution and quality. What’s more, they do not incorporate the physics underlying photosphere structure, and thus do not generate new knowledge about these features. We present work on an unsupervised machine learning approach to the problem of solar image segmentation, using a WNet model that produces its own definitions of features. Instead of relying on (or replicating) overly-simplistic definitions, this model creates a more robust, nuanced, and perhaps scientifically interesting framework for the identification of features within the solar photosphere.