Research on Instance Segmentation of Sunspot Groups Based on ConvNeXt-V2

Abstract Classification and segmentation of sunspot groups are valuable for analyzing the evolution of sunspot magnetic fields, which is closely associated with solar flare eruptions. This study presents a deep learning-based method tailored for the Mount Wilson magnetic classification and segmentation of sunspot groups. Utilizing data from the Helioseismic and Magnetic Imager collected from 2010 June to 2024 April, we created a labeled data set of 3916 composite images. The data set was split into training and testing sets in a 4:1 ratio to facilitate practical model training and evaluation. The study applies the ConvNeXt-V2 model, which integrating a Fully Convolutional Masked Autoencoder and a Global Response Normalization layer to enhance feature extraction capabilities and computational efficiency. The model achieves significant parameter reduction by leveraging the inverted residual bottleneck structure and depthwise separable convolutions. During training, a combination of Stochastic Gradient Descent and AdamW optimizers improves model accuracy and generalization ability. The model demonstrates strong performance in classification and segmentation tasks, attaining a mean [email protected] of 0.90 in classification and 0.88 in segmentation. Mean Precision exceeds 0.9, with mean Recall and F1-Score nearing 0.9 in classification task. Confusion matrix analysis reveals minor confusions between similar categories. Additionally, confidence heatmap analysis shows that smaller sunspots display lower segmentation confidence scores, highlighting challenges in small-target segmentation.