Classification of High-resolution Solar H{\alpha} Spectra using t-distributed Stochastic Neighbor Embedding

TL;DR

This study applies t-SNE, a machine learning technique, to classify high-resolution solar H-alpha spectra, effectively distinguishing different chromospheric features and aiding solar physics research.

Contribution

It demonstrates the use of t-SNE for nonlinear dimensionality reduction and classification of solar spectra, linking spectral clusters to chromospheric features with detailed parameter analysis.

Findings

t-SNE effectively separates quiet-Sun and plage spectra

Classification accuracy depends on t-SNE parameters and seeing conditions

Clusters correspond to distinct chromospheric features

Abstract

The H{\alpha} spectral line is a well-studied absorption line revealing properties of the highly structured and dynamic solar chromosphere. Typical features with distinct spectral signatures in H{\alpha} include filaments and prominences, bright active-region plages, superpenumbrae around sunspots, surges, flares, Ellerman bombs, filigree, and mottles and rosettes, among others. This study is based on high-spectral resolution H{\alpha} spectra obtained with the echelle spectrograph of the Vacuum Tower Telescope (VTT) located at Observatorio del Teide (ODT), Tenerife, Spain. The t-distributed Stochastic Neighbor Embedding (t-SNE) is a machine learning algorithm, which is used for nonlinear dimensionality reduction. In this application, it projects H{\alpha} spectra onto a two-dimensional map, where it becomes possible to classify the spectra according to results of Cloud Model (CM) inversions. The CM parameters optical depth, Doppler width, line-of-sight velocity, and source function describe properties of the cloud material. Initial results of t-SNE indicate its strong discriminatory power to separate quiet-Sun and plage profiles from those that are suitable for CM inversions. In addition, a detailed study of various t-SNE parameters is conducted, the impact of seeing conditions on the classification is assessed, results for various types of input data are compared, and the identified clusters are linked to chromospheric features. Although t-SNE proves to be efficient in clustering high-dimensional data, human inference is required at each step to interpret the results. This exploratory study provides a framework and ideas on how to tailor a classification scheme towards specific spectral data and science questions.