Abstract: Direction of arrival （DOA） estimation is a core technology in array signal processing that plays a crucial role in radar detection， wireless communication， and other fields. Traditional subspace algorithms suffer from significantly degraded angular estimation robustness under complex conditions such as a low signal-to-noise ratio （SNR） and few snapshots， which is attributed to the insufficient estimation accuracy of the covariance matrix. Meanwhile， existing deep-learning methods improve the feature extraction capabilities through data-driven approaches but fail to fully integrate the spatial topology of the array with the fine-grained phase and amplitude features of complex signals， resulting in limited performance in complex environments. To address these issues， this paper proposes a DOA estimation method based on a graph attention network （GAT）. The elements of a uniform linear array are modeled as graph nodes， with the real and imaginary parts of the autocorrelation values of the element covariance matrix serving as node features. The real and imaginary parts of the cross-correlation values of the covariance matrix between adjacent elements are used as edge features， thereby completely preserving the signal phase correlation and array topological information. The contribution weights of different elements in the angle estimation are autonomously learned to focus on key features by leveraging the GAT attention mechanism， resulting in the construction of an end-to-end angle regression model. Comparative experiments were conducted under multiple conditions with wide angle ranges， various SNRs， and different snapshot numbers to verify the performance of the proposed method compared with traditional subspace and mainstream deep-learning algorithms. The results demonstrated that the proposed method outperformed the comparison algorithms in angular estimation accuracy and robustness in complex scenarios， significantly enhancing the estimation performance in such environments with high precision advantages. This method provides a new technical approach for high-precision DOA estimation in complex electromagnetic environments， enriches the application concepts of graph neural networks in the field of array signal processing， and holds important guiding significance for engineering practice in the radar， communication， sonar， and other fields， with both theoretical research and practical application value.