A Hybrid System for Learning Sunspot Recognition and Classification

Sunspots observation and classification are important tasks for solar astronomers. The activity of sunspots can give clues to the timing of solar flares and the solar weather in general. This paper describes a hybrid system for automatic sunspot recognition and classification. The system uses a combination of image processing and machine learning techniques to process and classify sunspot groups from digital satellite images. Sunspot data are extracted from daily images of the solar disk captured by the NASA SOHO/MDI satellite. The classification scheme attempted was the seven-class Modified Zurich scheme. The main components of the hybrid system are: 1) the image processor, 2) the feature extractor, 3) the clusterer, 4) the classification learner. Furthermore, the paper compares two clustering algorithms: hierarchical average-link and a density-based DBSCAN and examines their usefulness in dealing with sunspot data. The aim is to create clusters that closely match "natural" sunspot groups. Clustering is an important step in the process as the classification performance is dependent on it. In previous papers we have shown that by combining clustering with classification the overall results can be substantially improved.