MCA-Net: a method based on deep learning for the classification of low-resolution stellar spectra

Classification of stellar spectra plays a crucial role in astronomical research, providing a wealth of valuable data and laying a solid foundation for humanity’s journey to explore the universe. However, existing studies often focus on the extraction of local features from stellar spectra, such as identifying peaks and troughs, which presents challenges for practical applications. While balancing stellar categories and their quantities, there remains room for improvement in classification accuracy. This study presents a new neural network, MCA-Net, which integrates convolutional neural networks with local feature extraction capabilities, long short-term memory networks adept at sequential data analysis, and attention mechanisms designed for long sequence mining. The aim is to effectively extract features from stellar spectral data and perform classification. The research utilizes stellar spectral data captured by the LAMOST telescope, encompassing classification tasks across three-class (F, G, K) and ten-class (A0, A5, F0, F5, G0, G5, K0, K5, M0, M5). Comparative experiments validated the effectiveness of the proposed method and network, demonstrating very high accuracy and F1-Score results in classification. The MCA-Net was tested using an independent test set, achieving an accuracy of 95.32% in the three-class classification task and an accuracy of 98.11% in the ten-class classification task.