Enhancing Hyperspectral Image Classification: A Hybrid CNN Approach with Spatial and Channel Attention Mechanisms

In the domain of hyperspectral imaging classification (HSIC) for remote sensing, recent advancements in deep learning have proven transformative. Convolutional neural networks have demonstrated immense potential in resolving these issues of HSIC, including insufficient labeled data, redundant spatial and spectral features, and overfitting. Traditional convolutional neural networks (CNNs) have effectively extracted spectral and spatial features, but 2D CNNs are limited in spatial modeling, while 3D CNNs alone struggle to distinguish spectral and spatial characteristics. Given that the accuracy of HSIC hinges on both spatial and spectral information, we proposed a hybrid-CNN model designed to mitigate the constraints of 2D and 3D CNNs. Our approach involves leveraging hybrid CNNs with spatial and channel attention (CA) mechanisms to address challenges like overfitting and model complexity. The proposed framework also improves generalization performance compared to 2D or 3D CNNs alone. Experiments were conducted on open datasets from Pavia University, Indian Pines, and Salinas to validate the proposed approach. The results demonstrate the efficacy of the hybrid CNN model with spatial and CA in consistently producing excellent classification results through thorough analyses with different deep-learning models.