Electroencephalography (EEG) is an effective tool for detecting Major Depressive Disorder (MDD). In cross-dataset EEG-based MDD detection, factors such as experimental paradigms, equipment variations, and individual differences result in notable disparities in domain distribution at both global and local levels. To tackle this challenge, this paper introduces an effective global-local distribution alignment (GLDA) strategy. Our method integrates dynamic graph convolution with a structure-adaptive updating mechanism to extract features. For global alignment, we employ Maximum Mean Discrepancy, which quantifies and minimizes the differences in statistical distributions between domains in high-dimensional spaces, ensuring a consistent feature representation across diverse datasets. For local alignment, Local Maximum Mean Discrepancy is utilized to refine this alignment by focusing on the distribution of subdomain samples that have the same labels, thus enhancing precision in heterogenous settings. This method effectively extracts potential correlations among EEG channels and mines spatial features, simultaneously narrowing the feature distribution gap across different domains at both global and local levels. Extensive validation was conducted on three established MDD datasets under both resting-state and task-state conditions. GLDA attained an average F1 score of 63.18% during the resting state and 63.80% while performing tasks, outperforming other methods in both scenarios.