| Summary: | With the rapid growth in astronomical spectra produced by large sky survey telescopes, traditional manual classification processes can no longer fulfill the requirements of precision and efficiency of spectral classification. There is an urgent need to employ machine learning approaches to conduct automated spectral classification tasks. Feature extraction is a critical step which has a great impact on any classification result. In this paper, a novel gradient-based method together with principal component analysis is proposed for the extraction of partial features of stellar spectra, that is, a feature vector indicating obvious local changes in data, which corresponds to the element line positions in the spectra. Furthermore, a general feature vector is utilized as an additional characteristic centering on the overall tendency of spectra, which can indicate stellar effective temperature. The two feature vectors and raw data are input into three neural networks, respectively, for training and each network votes for a predicted category of spectra. By selecting the class having the maximum votes, different types of spectra can be classified with high accuracy. The experimental results prove that a better performance can be achieved using the partial and general methods in this paper. The method could also be applied to other similar one-dimensional spectra, and the concepts proposed could ultimately expand the scope of machine learning application in astronomical spectral processing.
|
|---|
