<p>The temporal-frequency-spatial features of motor imagery electroencephalogram (EEG) signals provide comprehensive information for classification. However, these features also introduce significant redundancy and increase the feature dimension, complicating the classification task. Feature selection methods can identify subject-specific features and eliminate redundant information. Nevertheless, current feature selection approaches typically only remove features unrelated to the classification task labels, without addressing redundancy between features. To minimize feature redundancy, this paper proposes a two-stage framework combining feature selection and feature decorrelation. In the first stage, feature selection methods are employed to eliminate redundant and noisy information, thereby obtaining feature subsets closely related to the motor imagery task. In the second stage, the orthogonal transformation of principal component analysis (PCA) is utilized to remove correlations between features. Furthermore, kernel PCA (KPCA) and sparse PCA (SPCA) are designed to capture nonlinearity and sparsity among features, respectively. The optimal number of principal components for these PCA methods is determined using tenfold cross-validation, with classification accuracy as the evaluation criterion. The proposed method's effectiveness is validated on four motor imagery EEG datasets, achieving the highest average accuracies of 89.63%, 83.02%, 80.83%, and 79.40%, respectively. Experimental results demonstrate that the proposed methods significantly improve classification performance, with methods combining KPCA showing greater effectiveness. Additionally, the proposed methods enhance the separability of features, yielding good classification results even with simple classifiers like Fisher linear discriminant analysis.</p>