Classification of sample less hyperspectral images based on spatial-spectral fusion

Hyperspectral image classification plays a key role in the understanding of dynamic changes on the Earth's surface and in decision making in areas such as urban planning, agricultural land management, nature conservation, military intelligence, and security monitoring. Aiming at the problem of hyperspectral remote sensing image classification that limits its application due to the difficulty of obtaining a large number of high-quality labels, this study proposes a hyperspectral image classification model based on a few samples. The model extracts spatial and spectral features by mapping the original hyperspectral image data into spatial and spectral subspaces and utilizing a two-stream network to extract spatial and spectral features respectively. In the spatial subspace, a grouped convolution module and a spectral angle discrepancy measurement module are employed, aiming at capturing local spatial features and analyzing pixel similarity within the same region. While in the spectral subspace, the long-range dependency between spectra is effectively modeled based on the transformer model. Accurate classification of hyperspectral images is achieved by skillfully fusing the discriminative features of two-stream networks. Experimental results on two generalized public datasets show that the proposed method maintains the highest level of classification accuracy compared to other deep learning methods while using a small percentage of training samples.