A semi-empirical library of galaxy spectra for Gaia classification based on SDSS data and PÉGASE models

Aims. This paper is the third in a series implementing a classification system for Gaia observations of unresolved galaxies. The system makes use of template galaxy spectra in order to determine spectral classes and estimate intrinsic astrophysical parameters. In previous work we used synthetic galaxy spectra produced by PÉGASE.2 code to simulate Gaia observations and to test the performance of support vector machine (SVM) classifiers and parametrizers. Here we produce a semi-empirical library of galaxy spectra by fitting SDSS spectra with the previously produced synthetic libraries. We present (1) the semi-empirical library of galaxy spectra; (2) a comparison between the observed and synthetic spectra; and (3) first results of classification and parametrization experiments with simulated Gaia spectrophotometry of this library. Methods. We use χ2-fitting to fit SDSS galaxy spectra with the synthetic library in order to construct a semi-empirical library of galaxy spectra in which (1) the real spectra are extended by the synthetic ones in order to cover the full wavelength range of Gaia; and (2) astrophysical parameters are assigned to the SDSS spectra by the best fitting synthetic spectrum. The SVM models were trained with and applied to semi-empirical spectra. Tests were performed for the classification of spectral types and the estimation of the most significant galaxy parameters (in particular redshift, mass to light ratio and star formation history). Results. We produce a semi-empirical library of 33 670 galaxy spectra covering the wavelength range 250 to 1050 nm at a sampling of 1 nm or less. Using the results of the fitting of the SDSS spectra with our synthetic library, we investigate the range of the input model parameters that produces spectra which are in good agreement with observations. In general the results are very good for the majority of the synthetic spectra of early type, spiral and irregular galaxies, while they reveal problems in the models used to produce quenched star forming galaxies (QSFGs). The results of the SVM classification and regression models for this library are quite accurate for the prediction of the spectral type and the estimation of the redshift parameter, while they are quite poor for the cases of the most significant parameters used to produce the synthetic models (i.e. the star formation histories).