ObjectiveMotor Imagery (MI) is a key paradigm in Brain-Computer Interfaces (BCI) aimed at decoding motor intentions from EEG signals. However, accuracy remains challenging due to data limitations, noise, and non-stationarity. MethodThis paper introduces DFBRTS, a novel method leveraging Riemannian geometry and Cross-Frequency Coupling to improve MI-EEG decoding. DFBRTS filters EEG signals using a Dichotomous Filter Bank and employs Riemann Tangent Space Mapping for feature extraction in each sub-band. A lightweight convolutional neural network then performs further feature extraction and classification, supervised by cross-entropy and center loss. Experiments on the BCI competition IV 2a and OpenBMI datasets validate its effectiveness. ResultsDFBRTS outperforms state-of-the-art methods, achieving 78.16 % accuracy for four-class and 71.58 % for two-class classification, surpassing existing benchmarks. ConclusionDFBRTS significantly enhances MI-BCI decoding, promising improved interfaces for individuals with motor ‘‘the potential of DFBRTS to advance robust MI-BCI applications.