An important field of research in brain-computer interfaces is motor imagery recognition, which translates neural activity inputs from the brain into encoding outputs for intention control. The applications of deep learning algorithm have increased the precision of motor imagery recognition in recent years. Nevertheless, multi-channel electroencephalogram (EEG) inputs are typically treated as two-dimensional matrix signals in deep learning-based motor imagery analysis, which ignores spatial correlation information between distinct nodes. In order to overcome this problem, graph convolutional networks are used for node feature aggregation in motor imagery analysis, wherein spectral domain properties can be learned. Ultimately, fully connected layers are used to produce the categorization results. This approach learns temporal, frequency, and spectral domain information effectively, outperforming other methods with an accuracy of 80.9% and a kappa coefficient of 0.7 on Dataset 2a of BCI Competition IV. It offers a fresh viewpoint and method for BCI motor imagery recognition.