Sunspot Data Collection and Experimental Validation for McIntosh Classification

As an important basis for predicting solar activity, the McIntosh classification of sunspots is used by more and more international institutions and astronomical institutes because some categories are closely related to flare eruption. With the rapid increase in the amount of data, automatic McIntosh classification of sunspots has become an urgent need. Using the 720s-SHARP series data products provided by SDO/HMI and SRS files from NOAA as images and labels for McIntosh classification, this paper first augmented valid samples of a complete solar cycle (time span of 12 years) and cleaned data to establish the sunspot database newSharp on the basis of the Sharp database with only 7-year data. Secondly, in view of the characteristics of sunspot images, a series of preprocessing operations such as data allocation by active region number were taken, and proved its rationality and necessity. Finally, four classical classification neural network models in CNN were used to compare Sharp and newSharp for McIntosh classification experiments. The results show that compared with Sharp, newSharp not only has a significant increase in the amount of data, but also has better weighted F₁ score of most categories by augmenting valid samples and cleaning invalid samples. Besides, the weighted F₁ score of categories with a small number from newSharp even has achieved a breakthrough of 0. Over all, the weighted F₁ score of McIntosh-p improved the most, which greatly verifies the effectiveness of establishing a complete and reliable database and proves the rationality of using scientific and reasonable experimental methods. Thus it is able to better automatedly realize the end-to-end McIntosh classification tasks of sunspot images that are actually observed.