Sunspots Extraction in PMO Sunspot Drawings Based on Deep Learning

Abstract

Sunspot numbers and sunspot areas are the most fundamental indices of long-term solar activity levels and the solar magnetic dynamo. This paper presents a deep-learning method for segmenting the components of sunspots in the Purple Mountain Astronomical Observatory (PMO) historical hand drawings spanning from 1954 to 2011. A total of 44568 samples were labeled as the following four types to build the training set and the test set at a ratio of 9:1. They are (1) pores without penumbrae, (2) spots with penumbrae, (3) umbrae within spots, and (4) holes within spots. A Hybrid Task Cascade Region-based Convolutional Neural Networks (HTC RCNN) is adopted; it is designed as three cascade stages adapted to increasingly higher Intersection over Unions to obtain increasing detector quality. The features of sunspots are extracted and fused by a backbone combining residual network 50 and a feature pyramid network. After training the network using the training set, the method is tested by the test set. The precision, recall, and mean Average Precision are 0.90, 0.90, and 0.89, respectively, indicating that the method has a satisfying performance. The components of each sunspot are extracted separately, yielding the numbers and areas of pores, spots, umbrae and penumbrae. Detailed data of the PMO drawings from 1954 to 2011 are shared in public (http://61.166.157.71/HTCSD.html). It is another piece of the puzzle of long-term solar activity cycles and solar magnetic dynamo research.