Motor imagery electroencephalography (MI-EEG) decoding remains challenging due to low signal-to-noise ratio and complex temporal-spectral characteristics. This study aims to develop a robust deep learning framework for effective EEG representation learning. We propose EEG-DBNet, a dual-branch neural network that jointly models temporal dynamics and spectral representations of EEG signals. The model integrates local and global convolutional modules to enable multi-scale feature extraction, complementing the dual-branch design for multi-dimensional temporal-spectral representation learning. To validate robustness, experiments are conducted on two public datasets as well as a self-collected MI-EEG dataset acquired under controlled laboratory conditions. Experimental results show that EEG-DBNet achieves the best average performance on the two public benchmark datasets, BCI Competition IV-2a and IV-2b. On the self-collected CQUPT dataset, EEG-DBNet obtains competitive performance compared with representative baseline methods, suggesting its potential applicability to laboratory-acquired MI-EEG decoding. These results indicate that the proposed temporal-spectral dual-branch design is effective, while further validation on larger self-collected datasets is still needed. The proposed EEG-DBNet provides an effective solution for MI-EEG decoding with improved robustness. The inclusion of multiple datasets, particularly laboratory-acquired self-collected data, highlights its potential for practical brain-computer interface applications.