Dimensionality reduction on hyperspectral images: A comparative review based on artificial datas

In this research we address the problem of high-dimensional in hyperspectral images, which may contain rare /anomaly vectors introduced in the subspace observation that we wish to preserve. Linear techniques Principal Component Analysis(PCA), and non linear techniques Kernel PCA, Isomap, Multidimensional scaling (MDS), Local Tangent Space Alignment (LTSA), Diffusion maps, Sammon mapping, Symmetric Stochastic Neighbor Embedding (SymSNE), Stochastic Neighbor Embedding (SNE), Locally Linear Embedding(LLE), Locality Preserving Projection(LPP), Neighborhood Preserving embedding (NPE), Linear Local Tangent Space Alignment (LLTSA) was presented. Classical approaches criterion based on the norm l_d, derivative spectral, nearest neighbors and quality criteria are used for obtaining a good preservation of these vectors in the reduction dimension. We have observed from the results obtained that Sammon and Isomap are less sensitive to these rare vectors compared to the other presented methods.