Hybrid Dimensionality Reduction Technique for Hyperspectral Images Using Random Projection and Manifold Learning

Abstract

Hyperspectral images (HSI) are contiguous band images having hundreds of bands. However, most of the bands are redundant and irrelevant. Curse of dimensionality is a significant problem in hyperspectral image analysis. The band extraction technique is one of the dimensionality reduction (DR) method applicable in HSI. Linear dimensionality reduction techniques fail for hyperspectral images due to its nonlinearity nature. Nonlinear reduction techniques are computationally complex. Therefore this paper introduces a hybrid dimensionality reduction technique for band extraction in hyperspectral images. It is a combination of linear random projection (RP) and nonlinear technique. The random projection method reduces the dimensionality of hyperspectral images linearly using either Gaussian or Sparse distribution matrix. Sparse random projection (SRP) is computationally less complex. This reduced image is fed into a nonlinear technique and performs band extraction in minimal computational time and maximum classification accuracy. For experimental analysis of the proposed method, the hybrid technique is compared with Kernel PCA (KPCA) using different random matrix and found a promising improvement in results for their hybrid models in minimum computation time than classic nonlinear technique.