Deep Learning for Spectral Typing of Stars: A CNN Framework on Sloan Digital Sky Survey Observations

We present a Convolutional Neural Network (CNN) model for automated stellar classification using data from the Sloan Digital Sky Survey (SDSS) DR17. The proposed model achieved a precision of 0.94, recall of 0.91, and F1-score of 0.92, surpassing traditional approaches such as Random Forest (F1-score 0.83). The CNN demonstrated robustness under noisy conditions (F1-score 0.89) and achieved superior accuracy for M-type stars (F1-score 0.96), while minor confusion between G- and K-types was observed due to color similarities. Feature attribution using SHAP values revealed that g−r and r−i indices were the most influential, reflecting their physical correlation with stellar temperature. The model’s generalization was confirmed through 5-fold cross-validation (mean F1-score 0.91 ± 0.02) and high AUC scores (0.90–0.98). These results highlight the potential of CNN-based methods for scalable, explainable, and reliable star classification, paving the way for integration with Gaia, WISE, and LSST datasets for advanced astronomical analyses.