To accurately capture and address the multi-dimensional feature variations in cross-subject motor imagery electroencephalogram (MI-EEG), this paper proposes a time-frequency transform and Riemannian manifold based domain adaptation network (TFRMDANet) in a high-dimensional brain source space. Source imaging technology was employed to reconstruct neural electrical activity and compute regional cortical dipoles, and the multi-subband time-frequency feature data were constructed via wavelet transform. The two-stage multi-branch time–frequency–spatial feature extractor with squeeze-and-excitation (SE) modules was designed to extract local features and cross-scale global features from each subband, and the channel attention and multi-scale feature fusion were introduced simultaneously for feature enhancement. A Riemannian manifold embedding-based structural feature extractor was used to capture high-order discriminative features, while adversarial training promoted domain-invariant feature learning. Experiments on public BCI Competition IV dataset 2a and High-Gamma dataset showed that TFRMDANet achieved classification accuracies of 77.82% and 90.47%, with Kappa values of 0.704 and 0.826, and F1-scores of 0.780 and 0.905, respectively. The results demonstrate that cortical dipoles provide accurate time–frequency representations of MI features, and the unique multi-branch architecture along with its strong time–frequency–spatial–structural feature extraction capability enables effective domain adaptation enhancement in brain source space.