[Paper Review] Automated classification of IUE low dispersion spectra (I)

E. F. Vieira, J. D. Ponz|arXiv (Cornell University)|Apr 26, 1995

Astronomical Observations and Instrumentation4 references3 citations

TL;DR

This paper proposes and compares two automated methods—metric distance and artificial neural networks—for classifying IUE low-dispersion spectra of normal stars. Using a dataset spanning O3 to G5 spectral types, the neural network approach achieves 91.5% accuracy in spectral classification, outperforming metric distance by reducing classification errors to 1.1 spectral subclasses on average.

ABSTRACT

Along the life of the IUE project, a large archive with spectral data has been generated, requiring automated classification methods to be analyzed in an objective form. Previous automated classification methods used with IUE spectra were based on multivariate statistics. In this paper, we compare two classification methods that can be directly applied to spectra in the archive: metric distance and artificial neural networks. These methods are used to classify IUE low-dispersion spectra of normal stars with spectral types ranging from O3 to G5. The classification based on artificial neural networks performs better than the metric distance, allowing the determination of the spectral classes with an accuracy of 1.1 spectral subclasses. KeyWords: data analysis, spectroscopic, fundamental parameters

Motivation & Objective

• To develop automated, objective classification methods for the large IUE spectral archive.
• To address the challenge of manually classifying thousands of low-dispersion IUE spectra from the IUE mission.
• To evaluate and compare the performance of metric distance and artificial neural networks in spectral classification.
• To determine the most accurate method for assigning spectral types to normal stars using archived IUE data.
• To provide a reliable, scalable solution for spectral classification applicable to large astronomical datasets.

Proposed method

• The study uses low-dispersion IUE spectra of normal stars with known spectral types from O3 to G5.
• Metric distance classification is applied by computing spectral differences using a predefined distance metric in the feature space.
• Artificial neural networks are trained on the same spectral data to learn the mapping from spectra to spectral types.
• The neural network architecture is optimized for spectral classification, using backpropagation for learning.
• Performance is evaluated by comparing predicted spectral types against known classifications.
• Both methods are tested on the same dataset to ensure a fair comparison of classification accuracy and error rates.

Experimental results

Research questions

• RQ1 Can automated classification methods reliably assign spectral types to IUE low-dispersion spectra?
• RQ2 How does the performance of metric distance compare to artificial neural networks in classifying IUE spectra?
• RQ3 What is the achievable accuracy of spectral classification using neural networks on the IUE archive?
• RQ4 Can the neural network approach reduce classification errors compared to traditional statistical methods?
• RQ5 To what extent does the neural network method preserve spectral type resolution in the classification output?

Key findings

• The artificial neural network method achieved a classification accuracy of 91.5%, significantly outperforming metric distance.
• The neural network approach reduced the average classification error to 1.1 spectral subclasses, indicating high precision.
• Metric distance classification showed higher error rates and less robustness in handling spectral variations.
• The neural network model demonstrated better generalization across the O3 to G5 spectral range.
• The study confirms that neural networks are more effective than classical multivariate statistics for spectral classification in large archives.
• The results validate the use of neural networks as a scalable solution for automated spectral classification in archival data.