Solar thematic maps for space weather operations

Abstract

Thematic maps are arrays of labels, or “themes,” associated with discrete locations in space and time. Borrowing heavily from the terrestrial remote sensing discipline, a numerical technique based on Bayes' theorem captures operational expertise in the form of trained theme statistics, then uses this to automatically assign labels to solar image pixels. Ultimately, regular thematic maps of the solar corona will be generated from high‐cadence, high‐resolution SUVI images, the solar ultraviolet imager slated to fly on NOAA's next‐generation GOES‐R series of satellites starting ∼2016. These thematic maps will not only provide quicker, more consistent synoptic views of the sun for space weather forecasters, but digital thematic pixel masks (e.g., coronal hole, active region, flare, etc.), necessary for a new generation of operational solar data products, will be generated. This paper presents the mathematical underpinnings of our thematic mapper, as well as some practical algorithmic considerations. Then, using images from the Solar Dynamics Observatory (SDO) Advanced Imaging Array (AIA) as test data, it presents results from validation experiments designed to ascertain the robustness of the technique with respect to differing expert opinions and changing solar conditions.

Supervised classification of solar features using prior information

“…This allows estimating the PDF p(I(x)|C i ) for each class C i . Often, a parametric PDF is assumed and its parameters are estimated via ML (Turmon et al 2002;Dudok de Wit 2006;Rigler et al 2012;. In this work, motivated by the noise statistics analysis of Section 3.1, and in order to be as general as possible, we estimate these PDFs using a nonparametric kernel density estimator.…”

“…In our case, the test dataset is constituted of new manually segmented images. While Rigler et al (2012) studied the effect of having different experts specify the training and the test segmentation masks, in this work the test masks were created by the same person who made the initial training mask in order to avoid personal biases from influencing the results. Note that they still do not provide an ideal ground truth since these manual segmentations may contain errors themselves.…”

“…This effect may hinder the segmentation process and hence needs to be removed. In Rigler et al (2012), this is solved by assuming a constant height above the solar surface and computing the path length through this volume for the angle of view corresponding to each pixel. The corresponding image is then used as a pseudo-channel in a multi-channel procedure and accounts for limb brightening effects.…”

“…This is possible thanks to the ground truth' segmentation provided at training time. Previous work on supervised classification in solar images has used a parametric PDF in the form of a (possibly multivariate) Gaussian (Dudok de Wit 2006;Rigler et al 2012; or mixture of Gaussians (Turmon et al 2002) for such estimation. The Bayesian framework allows introducing prior information, e.g.…”

“…The Bayesian framework allows introducing prior information, e.g. in the form of enforcing spatial smoothness (Turmon et al 2002;Rigler et al 2012).…”