Abstract: Hyperspectral images classification is a research hotspot in the remote sensing field. The key point is to improve the classification accuracy by taking the advantage of spectral-spatial features of hyperspectral images, which fuse the spectral information and spatial information of each pixel in the hyperspectral image simultaneously. This paper employed multiview subspace learning for feature reduction with the problems of high feature dimension and redundant information of hyperspectral images, and proposed a graph regularized multiview marginal discriminant projection (GMMDP) algorithm. The multiview feature reduction algorithm took the spectral features of each pixels as a view and spatial features as another view, then searched the optimal discriminant common subspace by optimizing the projection direction of each view. Experiments on the open dataset showed that multiview learning had a significant advantage in spectral-spatial classification of hyperspectral images. Among all multiview dimensionality reduction algorithms, our GMMDP had the highest classification accuracy.

Key words: multiview learning, discriminant reduction, hyperspectral classification, spectral-spatial fusion