Decoding speech imagery (SI) from electroencephalogram (EEG) signals holds immense promise for individuals with severe speech production deficits. However, existing methods struggle with harnessing the diverse information inherent in EEG signals and mitigating their non-stationary nature. To address these limitations, this paper proposes a dual-branch riemannian network (DBRNet), which integrates parallel Riemannian manifold structures into a deep learning framework to simultaneously capture time-frequency and spatial features of EEG data. First, feature extraction decomposes the input sequence into feature maps representing time-frequency and spatial information. Next, Riemannian structures in the manifold space for each feature type are extracted independently. Then, multi-stage CNN-Transformer modules with feature fusion are employed for deep feature extraction and integration across stages, culminating in the final decoding results. Experimental results on the ASU dataset demonstrate that DBRNet significantly outperforms baseline methods in SI-EEG decoding, underscoring its potential as an advanced approach for this domain.