CNNs can effectively extract features with low computational costs, achieving significant progress in hyperspectral image classification. However, due to the limited receptive field of CNNs, they have difficulty in capturing the multi-scale structural and global contextual information. Moreover, the class imbalance in hyperspectral images often causes the model to focus disproportionately on certain spectral bands, thereby reducing the average accuracy. To address these challenges, a method called the Cascaded Spatial-Frequency Convolutional Network (CSFCNet) was proposed for hyperspectral image classification. It integrates rich spatial-domain information and frequency-domain information by jointly modeling both domains. Specifically, a Dual Spatial Fourier Convolution (DSF-Conv) module was proposed to project feature maps into parallel spatial and frequency representations. In the Spatial pathway, input features are grouped and processed with multi-scale convolutions to extract hierarchical structures; in the Fourier pathway, frequency-domain convolutions can aggregate the global context. Subsequently, a group-cascaded structure connects the DSF-Conv modules with residual connections, alleviating the class imbalance problem by promoting more balanced contributions from different spectral components. Additionally, we introduce a Lightweight Local Attention module to enhance the feature discrimination. Furthermore, experiments on three datasets achieved competitive accuracies, demonstrating the effectiveness of CSFCNet. Ablation studies further verify the effectiveness of the core components within the network.
Jiang et al. (Thu,) studied this question.