A Machine Learning Approach towards Segmentation and Analysis of Solar Filaments from Kodaikanal Solar Observatory Hand-drawn Archive

Abstract:

Kodaikanal Solar Observatory (KoSO) is identified as an important source of historical solar data offering century-long simultaneous observation at different wavelengths such as white light, Ca II K, and Hα since early 1900’s. These observations were originally recorded in photographic films/plates and are currently available publicly in digitised form. In addition, KoSO has stored daily ‘sun-charts’ that constitute composite data from all wavelengths in the form of coloured drawings made on Stonyhurst latitude-longitude grid. Those present solar features such as sunspots, plages, filaments and prominences. These sun-charts are digitized using industry level scanner and are stored in digital (‘.tif’) format. For our analysis, we have used data for a period of 2 solar cycles (1954-1976) and focused on detection of filaments that align with polarity inversion lines and trace polar magnetic field reversal. Even though the filaments were marked with unique color in sun-charts by manual operators, the red (R), green (G) and blue planes (B) values of scanned images were not constant. This inspired us to implement an unsupervised machine learning technique called ‘k-means clustering’ to extract filaments with [R, G, B] as feature vector. To assign physical coordinates to filaments we implemented a novel automated technique to detect solar disc. We extracted different parameters such as centroid latitude, tilt angle, length, area from individual filaments and could produce physically consistent results. Our study showed increase in filament length and area with latitude, and clearly depicted poleward migration with dominance of a filament tilt sign. We further created Carrington maps of filaments and found good match while comparing with hand-drawn maps of Meudon database for overlapping rotations. Our study thus establishes consistency of KoSO hand-drawn archive and complements the existing hand-drawn historical Synoptic series with full-disc filament data.