Spectral-spatial hyperspectral image classification based on capsule network with limited training samples

For hyperspectral images, the classification problem of limited training samples is an enormous challenge, and the lack of training samples is an essential factor that affects the classification accuracy. Making full use of the rich spectral and spatial information contained in hyperspectral images can achieve a high-precision classification of hyperspectral images with limited training samples. This paper proposed a new spectral-spatial feature extraction method and constructed a new classification framework for limited training samples. In our work, deep reinforcement learning is used to process the spectral features of hyperspectral images, and then the selected bands with the most information are used to extract the spatial features by extended morphological profiles. Finally, the hyperspectral image after feature extraction is classified by the capsule network, which is more effective at exploiting the relationships between hyperspectral imaging features in the spectral-spatial domain. We conducted classification experiments on five well-known hyperspectral image data sets. The experimental results reveal that the proposed method can better extract the spectral and spatial features in hyperspectral images, achieve higher classification accuracy with limited training samples, and reduce computational and time complexity.