Joint spatial multiscale and frequency domain features for hyperspectral image classification

Traditional hyperspectral classification usually focuses on spectral features, spatial feature extraction and classifier design. With the wide application of deep neural networks, the joint application of shallow features and deep features makes it possible to classify hyperspectral images at a finer level. In this paper, a hyperspectral image classification method with joint spatial multiscale and frequency domain features (SMFDF) is proposed. The method introduces spatial multiscale feature module (SMFM) and frequency domain-channel attention module (FDCAM), aiming to solve the problems of insufficient feature extraction and lack of frequency domain information in traditional methods. SMFM utilizes multi-layer pooling operations to deeply mine spatial features at different scales, while FDCAM combines channel attention and frequency domain feature extraction to improve feature expression by guiding the frequency domain information selection through the attention mechanism. Experimental results show that the proposed method can significantly improve the classification accuracy in hyperspectral image classification tasks.