Automatic Classification of Stellar Spectra
Iciar Carricajo, Minia Manteiga, A. Rodríguez, Carlos Dafonte · 2004
Abstract: We propose and discuss the application of Artificial Intelligence techniques to the classification of stellar spectra. Two types of systems are considered, knowledge-based systems (Expert Systems) and different classes of neural networks. After analysing and comparing the performance of both systems in the classification of stellar spectra, we reach the conclusion that neural networks are more adequate to determine the spectral types and luminosity of stars, whereas knowledge-based systems are more performative in determining global temperatures. In order to determine the best approach to the classification of each spectrum type, we describe and analyse the performance and results of various neural networks models. Backpropagation networks, self-organising maps and RBF networks in particular were designed and tested, through the implementation of different topologies, to obtain the global classification, spectral type and luminosity of stars. The best networks reached a success rate of approximately 97 % for a sample of 100 testing spectra. The morphological analysis algorithms that were developed in the knowledgebased systems are used to extract and measure spectral features, and to obtain the input patterns of the neural networks. Some networks were trained with this parameterisation, others with flux values of specific spectral zones; it was the first strategy that resulted in a better performance. Our approach is focused on the integration of several techniques in a unique hybrid system. In particular, signal processing, expert systems, fuzzy logic and artificial neural networks are integrated by means of a relational database, which allows us to structure the collected astronomical data and to contrast the results of the different classification methods. In addition, we designed several models of artificial neural networks that were trained with synthetic spectra, and included them as an alternative classification method. 153 Iciar Carricajo et al.