Electroencephalography (EEG) is a non-invasive technique widely used in neuroscience and brain-computer interfaces (BCI) due to its high temporal resolution. In motor imagery EEG (MI-EEG) tasks, EEG signals reflect movement-related brain activity, making them ideal for BCI control. However, the non-stationary nature of MI-EEG signals poses significant challenges for classification, as frequency characteristics vary across tasks and individuals. Traditional preprocessing methods, such as bandpass filtering and standardization, may struggle to adapt to these variations, potentially limiting classification performance. To address this issue, this study introduces EA-EEG, an improved MI-EEG classification model that incorporates whitening as a preprocessing step to reduce channel correlation and enhance the model feature extraction ability. EA-EEG further leverages a multi-scale pooling strategy, combining convolutional networks and root mean square pooling to extract key spatial and temporal features, and applies prototype-based classification to improve MI-EEG classification performance. Experiments on the BCI4-2A and BCI4-2B datasets demonstrate that EA-EEG achieves state-of-the-art performance, with 85.33% accuracy (Kappa = 0.804) on BCI4-2A and 88.05% accuracy (Kappa = 0.761) on BCI4-2B, surpassing existing approaches. These results confirm EA-EEG's effectiveness in handling non-stationary MI-EEG signals, demonstrating its potential for robust BCI applications, including rehabilitation, prosthetic control, and cognitive monitoring.