Automated sunspot classification using machine learning techniques

Abstract

Solar observation is an active area of research intriguing many astronomers. Solar images are supplied by observatories. Image processing and the classification of sunspots in particular is a labour intensive, time consuming job involving astronomers, computer scientists and military personnel [NOAA/USAF procedures manual) This thesis aims to help astronomers and astrophysicists in solar weather forecasting by developing a tool that automates sunspot classification. Solar features are the drivers for solar activity. A solar flare is a solar activity with a wide range of precursors. Sunspots as a solar feature is a well studied precursor. Using image processing and machine learning techniques our system attempts sunspot classification under the McIntosh Classification system. Pairs of images (continuum and magnetogram images) are processed to extract sunspot. These are then clustered are features are extracted. Finally groups are classified. The DBSCAN clustering algorithm, Neural networks and fuzzy logic controllers are used in the process. Evaluation is conducted against experts data and results from a similar system. Results are satisfactory when seen in the light of the real situation of the problem being tackled. Incorporating this system into a fully fledged system that takes into account other solar flare precursors is the ultimate goal of our study.