000003351 001__ 3351
000003351 005__ 20180111115047.0
000003351 037__ $$aCTALK-2018-0009
000003351 100__ $$aDelouille, Véronique  
000003351 245__ $$aSupervised and non-supervised classification in solar physics using advanced techniques
000003351 260__ $$c2017
000003351 269__ $$c2017-09-26
000003351 520__ $$aAdvances in statistical signal processing allow us to gaze at solar data from a new perspective, and to make better predictions. For example, the problem of clustering active regions and sunspots from magnetogram and continuum data can be looked at with image patch analysis  and matrix factorization lenses.  Such method provides a characterization of fine scale structures encoded e.g. by localized gradients, or locally smooth areas. The resulting clusterings are related to large scale descriptors of an active region such as its size, local magnetic field distribution, and complexity as measured by Mount Wilson classification.  Supervised classification in the presence of an imbalanced dataset is another example where recent advances bring added accuracy. I will illustrate this on the problem of separating filaments (FL) from coronal holes (CH) using a labelled dataset of features, where the FL/CH proportion in the observed sample is 6/94. Various strategies for dealing with imbalance will be discussed. This is a generic problem, that may also appear e.g. when one want to distinguish flare-productive from more quiet active regions.
000003351 594__ $$aNO
000003351 700__ $$aMoon, Kevin
000003351 700__ $$aHero, Alfred O.
000003351 700__ $$aHofmeister, Stefan
000003351 700__ $$aReiss, Martin 
000003351 700__ $$aVeronig, Astrid
000003351 773__ $$tSpace Weather: a Multi-Disciplinary Approach, Lorentz Center workshop, Leiden, NL
000003351 8560_ $$fveronique.delouille@observatoire.be
000003351 906__ $$aInvited
000003351 980__ $$aCTALKINVI