Land cover classification is one of the most important applications of polarimetric SAR images, especially in urban areas. There are numerous features that can be extracted from these images, hence feature selection plays an important role in PolSAR image classification. In this study, three main steps are used to address this task: (1) feature extraction in the form of three categories, namely original data features, decomposition features, and SAR discriminators; (2) feature selection in the framework of the single and multi-objective optimization; and (3) image classification using the best subset of features. In single objective methods, we employ genetic algorithms (GAs) and support vector machines (SVMs) or multi-layer perceptron (MLP) neural network in order to maximize classification accuracy. Then a new method is proposed to perform an efficient land cover classification of the San Francisco Bay urban area based on the multi-objective optimization approach. The objectives are to minimize the error of classification and the number of selected PolSAR parameters. The experimental results on Radarsat-2 fine-quad data show that the proposed method outperforms the single objective approaches tested against it, while saving computational complexity. Finally, we show that the our method has a better performance than the SVM with full set of features and the Wishart classifier which is based on the covariance matrix.