Stellar classification and imaging
Pro Research Analysisby 
Searched over 200M research papers
Advances in Stellar Classification Using Imaging and Machine Learning
Stellar classification has evolved rapidly with the growth of large-scale sky surveys, which provide vast amounts of photometric and imaging data. Modern approaches increasingly rely on machine learning and deep learning to classify stars and stellar clusters, especially when spectroscopic data is limited or unavailable.
Photometric Imaging and Deep Learning for Stellar Classification
Recent research demonstrates that photometric images, especially those synthesized from multiple bands, can be effectively used for stellar classification. Vision Transformer (ViT) models, such as stellar-ViT, have shown superior accuracy compared to traditional convolutional neural networks (CNNs) and previous state-of-the-art models. For example, using RGB images from SDSS photometric data, stellar-ViT achieved an accuracy of 0.839, which increased to 0.863 with the inclusion of additional bands, highlighting the value of multi-band imaging for classification tasks . Similarly, CNN-based models like SCNet have been used to classify millions of SDSS stars into seven main spectral classes, achieving high accuracy (0.861) and enabling the creation of large star catalogues without spectra .
Single-Band and Multimodal Imaging Approaches
Even single-band imaging can be leveraged for stellar classification. Machine learning methods using principal component analysis (PCA) and artificial neural networks (ANNs) have successfully classified stars into spectral subclasses based solely on the shape of their diffraction patterns in single broad-band images, with typical errors of only half a spectral class . For more specialized tasks, such as M-type star classification, multimodal networks that fuse both spectral and photometric image data using Transformer architectures have achieved high F1-scores (up to 95.65%), demonstrating the benefit of combining multiple data types .
Image-Based Feature Extraction and Classification Algorithms
Advanced feature extraction techniques further enhance classification accuracy. Methods like the DRsm algorithm use DenseNet and ResNet models to extract features from synthetic RGB images of stellar spectra, achieving classification accuracies as high as 0.96, outperforming several other machine learning approaches . Capsule networks, which preserve hierarchical relationships in image data, have also been shown to improve classification performance for specific spectral types .
Machine Learning for Stellar Blend and Cluster Identification
Imaging challenges such as stellar blends—where multiple stars appear as a single source—are being addressed with computationally efficient models like Gaussian Processes (MuyGPs), which can distinguish blends from single stars with high accuracy (83.8%) and provide confidence estimates for human-assisted labeling . For the identification and classification of stellar clusters, pipelines developed for large imaging surveys (e.g., PHANGS-HST) use morphological parameters like the Multiple Concentration Index (MCI) and integrate machine learning to distinguish clusters from stars, achieving high selection purity and enabling deeper, more complete catalogues 67.
Automated Classification in Crowded and Complex Fields
In crowded regions such as the Galactic Centre, random forest classifiers trained on photometric data have proven effective for distinguishing between early- and late-type stars, matching the performance of more complex models and requiring minimal calibration . These approaches are scalable and adaptable as more training data becomes available.
Conclusion
Stellar classification and imaging have been transformed by the integration of machine learning and deep learning techniques. By leveraging photometric images, advanced feature extraction, and multimodal data fusion, researchers can classify stars and stellar clusters with high accuracy, even in the absence of spectroscopic data. These methods are scalable, robust, and adaptable to the vast and growing datasets produced by modern sky surveys, paving the way for more comprehensive and automated stellar catalogues.
Sources and full results
Most relevant research papers on this topic