In recent years, studies on the automatic estimation of modal parameters have grown significantly. In this research, one of the algorithms for automatic estimation of modal parameters has been improved. The algorithm is formed by five steps. The algorithm begins with selecting a high model order. Then, it finds the model order close to the real model order and resamples the dataset by the use of statistical methods and the obtained model order. In the next step, it clusters the obtained modes. Then it removes the noise clusters from the obtained clusters. Finally, using three features and machine learning algorithms, it places the remaining clusters in two physical and noise categories. In this study, 13 new features are introduced to improve this algorithm. Using the feature selection algorithms, these 13 features and the three main features of the algorithm are compared and more efficient features are introduced. To compare these features, PCA feature selection algorithms, Laplacian score, infinite feature selection, and improved distance evaluation method are used. The PCA method is in the category of dimensions reduction methods. Laplacian score and the infinite feature selection methods are in the category of unsupervised methods and the improved distance evaluation method is in the category of supervised methods. Utilizing these methods leads to obtain better performance features. The improved algorithm is implemented on five case studies. Keywords: Suace based identification, Modal parameters estimation, Frequency response, Feature selection, Modal analysis, Machine learning