Deep Learning With Grouped Features for Spatial Spectral Classification of Hyperspectral Images

This letter presents a novel deep learning algorithm for feature extraction from the hyperspectral images. The proposed method takes advantage of the knowledge that the features of the spatial-spectral data naturally fall into an array of groups with respect to different spectral bands. Aiming to reduce the influence of redundant spectral bands adaptively using unlabeled hyperspectral data, we incorporate the group information in the training algorithm of the deep neural network via a regularized weight-decay process. Experiments over different benchmarks of hyperspectral images show that the proposed method provides competitive solution with the state-of-the-art approaches.