Morphological Classification of G-band Bright Points Based on Deep Learning

Abstract

The G -band bright points (GBPs) are driven by the granular convection flow, and appear with different morphologies in the quiet Sun, such as points, elongated chains, and knees. A point-like GBP represents a single slender flux tube, while the others represent the interaction of several slender flux tubes. This paper presents a deep-learning method, GBPs Mask Region-based Convolutional Neural Networks (GBP-MRCNN) for morphological classification of GBPs. A training set including 10,404 samples and a validation set including 2057 samples have been built. A total of 906 images as a testing set have been detected, segmented, and classified. They consist of 74,973 points (88.3%), 8660 elongated chains (10.2%), and 1239 knees (1.5%). GBP-MRCNN has achieved an efficient classification of GBPs with a mean average precision of 0.93. For points, elongated chains, and knees, the quiet Sun fractional area average values are 0.47%, 0.14%, and 0.03%, respectively. Correspondingly, the mean areas are (2.02 ± 0.69) × 10 4 , (6.17 ± 1.71) × 10 4 and (6.99 ± 1.74) × 10 4 km 2 , respectively. The maximum intensity contrasts are 1.05 ± 0.12, 1.18 ± 0.17, and 1.36 ± 0.17, respectively. The mean eccentricities of points and elongated chains are 0.48 ± 0.23 and 0.89 ± 0.01, respectively. We adopt constituent points to describe how many points constitute a nonpoint-like GBP. The mean constituent points of elongated chains and knees are estimated as 2.77 ± 0.57 and 3.06 ± 0.65, respectively. The maximum length of the elongated chains can extend to over 1000 km, which corresponds to over eight single slender flux tubes. The characteristics of GBPs with different morphologies can be explained by the interaction between magnetic flux tubes and photospheric convection, and then help to build improved magnetic flux tube models.