To address the challenge of insufficient capture of learnable functional connectivity and discriminative spatial-temporal features in motor imagery EEG (MI-EEG) decoding, this paper proposes a Spatial-Temporal Graph Convolutional Attentional Network (STGCAN-MI). The method first explicitly models the functional connectivity and temporal dependencies among electrodes through a learnable dynamic graph convolution module. It then introduces a lightweight parallel spatial-temporal attention module to perform adaptive recalibration of feature responses across channel and time dimensions, thereby enhancing critical discriminative information and suppressing irrelevant noise. Experiments on the BCI Competition IV dataset 2a (four-class) and 2b (binary-class) show that STGCAN-MI achieves average best classification accuracies of 83.17% and 87.50%, respectively, outperforming several representative baseline models. Ablation studies and visualization analysis validate the effectiveness of each core component. This research provides a new approach for developing high-performance decoding methods for motor imagery EEG.