Abstract: To address the dramatic decline in direction-of-arrival(DOA) estimation performance at low signal-to-noise ratio(SNR), this study introduces a convolutional neural network-based DOA estimation classification method. Trained on multi-channel data from real array manifolds, the proposed method directly estimates DOA, circumventing the performance limitations of spectral peak search methods. By employing the array covariance matrix's columns as under sampled noise linear measurements of spatial spectra, the method extracts the local upper triangular region of array signals as features, significantly reducing redundancy in data processing and enhancing the model's generalizability. Furthermore, by redefining the DOA estimation as a multi-label classification task and applying a binary classification strategy on an angle-space, estimation accuracy is improved. Simulation results demonstrate that, compared to traditional DOA estimation algorithms and fully connected neural network-based methods, our proposed algorithm achieves higher estimation accuracy, better robustness, and superior overall performance under low SNR, limited snapshots, and off-grid angles.