Motor imagery (MI) electroencephalography (EEG) signal decoding faces challenges such as low spatial resolution, effective feature extraction under signal-to-noise ratio constraints, and inter-individual variability. Although deep learning methods offer new ways to improve MI classification accuracy, they generally suffer from weak generalisation, high predictive volatility, and difficulties in balancing network decoding performance with computational resource consumption. To address these challenges, we propose GSoP-LENet, a multidimensional attention network for MI classification. This network integrates the Global Second-order Pooling (GSoP) mechanism with a multi-scale feature learning framework to enhance the accuracy and calibration efficiency of event-related potential (ERP)-assisted brain-computer interface (BCI) systems. The GSoP mechanism captures high-order dependencies between channels, while the design ensures efficient computation. We introduce two innovative modules: the GSoP-TCAB module, which combines multi-scale temporal convolution with channel covariance attention to capture frequency band-channel interactions; and the GSoP-STAF module, which models electrode spatial covariance matrices to extract dynamic ERP patterns. Experiments on BCI Competition IV-2a and 2b datasets demonstrate that GSoP-LENet achieves average classification accuracies of 80.58% and 83.21%, respectively. This model satisfies the need for high precision, and generalisation in non-invasive BCI systems.