Stellar spectral feature extraction and combination analysis for classification with ENN

This paper presents a novel stellar spectral classification method. Wavelet packet transform is adopted to extract continuums and absorptions in the spectra. Then a method for constructing combinatorial features is introduced, which is suitable for both temperature and luminosity classification. Finally both temperature and luminosity classes of the stars are determined using ensemble neural networks. Experiments with real world data show that the feature extraction process is efficient and the obtained correct classification rate is quite satisfying. The results also show that the ensemble neural networks give a better generalization than a single back propagation neural network.