Hyperspectral Image Classification via Improved Convolutional Neural Network with Spectral-Spatial Feature Fusion and Attention Mechanism

paper proposes a hyperspectral image classification method based on an improved Convolutional Neural Network (CNN) combined with an enhanced Squeeze-and-Excitation (SE) attention mechanism. The method effectively extracts spectral-spatial features through deep convolutional layers, while the enhanced SE module integrates multi-scale pooling and spatial attention to dynamically adjust the weights of spectral and spatial features, enhancing the model’s focus on critical information. To address the class imbalance problem in hyperspectral datasets, Gaussian noise injection and the Synthetic Minority Oversampling Technique (SMOTE) are introduced, improving the model’s robustness and generalization capability. Extensive experiments conducted on the Indian Pines, Pavia University, and Salinas datasets demonstrate that the proposed method achieves classification accuracies of $95.75 \%, 98.88 \%$, and 97.84%, with mean average precision (mAP) values of $89.41 \%, 99.33 \%$, and 93.12%, respectively. Compared to traditional CNNs and other classic methods, the proposed method shows significant advantages in classifying complex and similar categories, validating the effectiveness of combining spectral-spatial features with attention mechanisms.The proposed approach provides a new technical direction for hyperspectral image classification, offering substantial potential for improving classification accuracy and addressing class imbalance issues in high-dimensional data.