Learning Spatial-Spectral Features for Hyperspectral Image Classification

Combining spatial information with spectral information for classifying hyperspectral images can dramatically improve the performance. This paper proposes a simple but innovative framework to automatically generate spatial-spectral features for hyperspectral image classification. Two unsupervised learning methods-K-means and principal component analysis-are utilized to learn the spatial feature bases in each decorrelated spectral band. The spatial feature representations are extracted with these spatial feature bases. Then, spatial-spectral features are generated by concatenating the spatial feature representations in all/principal spectral bands. The experimental results indicate that the proposed method is flexible enough to generate rich spatial-spectral features and can outperform the other state-of-the-art methods.