Fully automated detection of solar activity manifested in spectral images of the solar disk holds significant scientific value for advancing solar physics research. This study formulates the task as a classification problem using localized images of the solar disk. We first construct a solar activity classification dataset derived from CHASE full-disk spectral images. This dataset comprises both single-channel H $\alpha $ images and multi-channel images spanning the H $\alpha $ and Fe I passbands of the CHASE data. These multi-channel data represent a novel resource, as prior studies have not explored solar activity recognition using dual-passband multi-channel data. Subsequently, we develop a classification model leveraging Residual Networks (ResNets), and by optimizing the network architecture and incorporating attention mechanisms, the model effectively captures visual features of solar activity from multi-channel spectral images. Furthermore, we introduce a strategy of spectral channel normalization and downsampling to improve the model’s classification accuracy and training efficiency. Comparative and ablation experiments confirm that the proposed model delivers robust classification accuracy and efficient inference performance on this dataset.
