Deep Learning for Active Region Classification: A Systematic Study from Convolutional Neural Networks to Vision Transformers

Abstract Solar active regions can significantly disrupt the Sun–Earth space environment, leading to severe space weather events such as solar flares or coronal mass ejections. Consequently, the automatic classification of active region groups is a crucial starting point for accurately and promptly predicting solar activity. This study presents our application of deep learning techniques to classify active region cutouts based on the Mount Wilson classification scheme. We explore the latest advantages in image classification architectures, ranging from convolutional neural networks to vision transformers, alongside modern training procedures, including on-the-fly data augmentations and transfer learning. We aim at evaluating the respective strengths and limitations of different neural network architectures in classifying solar active region cutouts. We observed that combining magnetogram and continuum image types enhances model performance by leveraging complementary features from diverse inputs. When considering only magnetograms, data-efficient image transformers achieve the best performance, suggesting that these models can better capture the spatial complexity of magnetograms. Models trained exclusively on continuum images exhibit overall lower performance, suggesting that continuum images, due to their more homogeneous nature, offer less spatial variability.