Virtual Dimensionality Analysis of Hyperspectral Imagery with Noise being Constrained

Abstract:

In dimensionality reduction process of hyperspectral data, intrinsic dimension is normally characterized by virtual dimension. Classic algorithm mainly uses hypothesis-testing criterion to set eigenvalue threshold and correspondingly obtains virtual dimension. But under strong noises, it may not estimate very well. A noise constrained virtual dimension (NCVD) analysis method of hyperspectral imagery is proposed in this paper. It decreases the computational complexity by the QR decomposing; improves the accuracy of the estimated dimension by adopting sliding noise detection window to filter the noise; synthesizes the least squares algorithm to modify threshold for reasonable results. The experimental results prove the feasibility and superiority of the proposed algorithm by using simulated and real data.