Filter Bank Regularized Riemannian Features Fusion Based on Dempster–Shafer Theory for Small Sample Motor Imagery Classification

ABSTRACT Over the past decade, the Riemannian geometry (RG)‐based method has proven to be an effective tool for motor imagery classification. However, its performance deteriorates drastically in small‐sample setting (SSS) scenarios and heavily depends on the frequency range of active EEG signals. Therefore, a novel regularized RG framework, named filter bank regularized common spatial pattern‐based tangent space (TS) and Dempster–Shafer fusion (fFBRCTS) method, is proposed in this study. In fFBRCTS, raw EEG signals are firstly divided into multiple subbands, and the most discriminative filter banks are chosen based on the classification accuracy of each subband. Then, TS features are estimated from each selected subband using a regularized common spatial pattern‐based TS model and optimized through a mutual information approach. Subsequently, a probabilistic support vector machine (PSVM) is employed for classification. Finally, a Dempster–Shafer theory‐based method is utilized to combine the classification results obtained under different regularization parameter pairs. Evaluation experiments conducted on the BCI Competition III dataset IVa validate its effectiveness, achieving an average accuracy of 87.0%, which surpasses other comparative methods. Moreover, an extended experiment performed on the BCI Competition III dataset I demonstrates its robustness, resulting in an average accuracy of 81.7%. These results confirm the efficacy of the proposed regularized RG framework in improving MI‐BCI performance under SSS conditions.