This paper considers the modal parameter identification problem for an LSS (Large Space Structure) from measured input-output data. The selection of setting parameters for identification, such as the sampling rate and identification model order, has great influence on the identification accuracy. This paper proposes a new identification method which uses decimated input-output data. The decimation consists of two parts, first the input-output data are filtered by low-pass filters, and then resampled by the lower sampling rate. The elastic modes can be identified accurately by proper choice of the cut-off frequency of the filter. In the proposed identification method, the model order is determined by the frequency domain criterion, that is, the flatness of the identification residual sequence. The effectiveness of the proposed identification method is shown by identification experiments which were carried out by using a ground-based test model.