DRsm: Star spectral classification algorithm based on multi-feature extraction

• Published in: New astronomy Vol. 116; p. 102349
• Main Authors: Yang, Jiaming, Tu, Liangping, Li, Jianxi, Miao, Jiawei
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.05.2025
• Subjects: DenseNet; LAMOST; ResNet; Stellar spectrum; LAMOST; DenseNet; Stellar spectrum; ResNet
• ISSN: 1384-1076
• DOI: 10.1016/j.newast.2024.102349

With the development of information technology, data-driven astronomical research has become a very popular subject. In view of the huge amount of spectral data from the sky, it is necessary to find suitable automatic processing methods to meet the needs of the time. Based on DenseNet model and ResNet model, DRsm (DenseNet ResNet SoftMax) algorithm is built in this paper, which realizes the automatic classification of stellar spectra. There are 6 steps to the DRsm algorithm: (1) Normalization processing: The Min–max normalization function is used to normalize the stellar spectrum to speed up the algorithm. (2) Denoising processing: The Ces algorithm is employed to denoise the stellar spectrum by reducing the photon noise that affects the spectral observations. (3) Composite RGB image: Three channels of an RGB image, corresponding to the gray image generated by the same spectrum. By superimposing the same spectrum, the effective distinguishing features of the stellar spectrum become more apparent and subsequent work is made easier. Here, we have normalized the continuous spectrum of the stellar spectrum, so that the content shown in the RGB image is basically the spectral line information of the star spectrum. At the same time, we analyze the feasibility of data conversion (synthetic RGB image) : using the main spectral line information of the star spectrum as a reference, we investigate whether the relevant pixel position of the synthesized RGB image contains these features. (4) Data enhancement: The Bottom-hat transformation (Top-hat transformation, contrast enhancement algorithm) is used to enhance the converted data, so that the main distinguishing features of the star spectrum are more obvious. (5) Feature extraction: The ResNet model and DenseNet models are used to extract features from stellar spectra, and the RGB image with a scale of 64 × 64 is extracted as a one-dimensional feature vector. (6) Automatic classification: The feature vector is then sent to the SoftMax module where it is automatically classified. The loss function used by the SoftMax module is ‘data set loss + regular term loss’. When the DRsm algorithm is used to automatically classify the spectra of A, B, dM, F, G, gM and K-type stars with R-band signal-to-noise ratio greater than 30, the classification accuracy is 0.96. The classification accuracy of this method is notably higher than that of the CNN(Convolutional Neural Networks)+Bayes, CNN+KNN, CNN+SVM, CNN+AdaBoost, and CNN+RF algorithms, which achieved accuracies of 0.862, 0.876, 0.894, 0.868, and 0.889, respectively. •Normalization processing: The Min-max normalization function is used to normalize the stellar spectrum to speed up the algorithm.•Denoising processing: The Ces algorithm is used to denoise the stellar spectrum to remove the photon noise doped by the stellar spectrum.•Composite RGB image: Three channels of an RGB image, corresponding to the gray image generated by the same spectrum. The superposition of the same spectrum makes the effective distinguishing features of the stellar spectrum more obvious and makes subsequent work easier. Here, we have normalized the continuous spectrum of the stellar spectrum, so that the content shown in the RGB image is basically the spectral line information of the stellar spectrum. At the same time, we analyze the feasibility of data conversion (synthetic RGB image): using the main spectral line information of the stellar spectrum as a reference, to investigate whether the relevant pixel position of the synthesized RGB image contains these features.•Data enhancement: The Bottom-hat transformation (Top-hat transformation, contrast enhancement algorithm) is used to enhance the converted data, so that the main distinguishing features of the stellar spectrum are more obvious.•Feature extraction: ResNet model and DenseNet model are used to extract features from stellar spectra, and the RGB image with a scale of 64×64 is extracted as a one-dimensional feature vector.•Automatic classification: The feature vector is sent to the SoftMax module for automatic classification.