Spectral target detection using a physical model and a manifold learning technique

摘要

Identification of materials from calibrated radiance data collected by an airborne imaging spectrometer depends strongly on the atmospheric and illumination conditions at the time of collection. This paper presents a methodology for identifying material spectra using the assumption that each unique material class forms a lower-dimensional manifold (surface) in the higher-dimensional spectral radiance space and that all image spectra reside on, or near, these theoretic manifolds. Using a physical model, a manifold characteristic of the target material exposed to varying illumination and atmospheric conditions is formed. A graph-based model is then applied to the radiance data to capture the intricate structure of each material manifold followed by the application of the commute time distance (CTD) transformation to separate the target manifold from the background. Detection algorithms are than applied in the CTD subspace. This nonlinear transformation is based on a Markov-chain model of a random walk on a graph and is derived from an eigendecomposition of the pseudoinverse of the graph Laplacian matrix. This paper discusses the properties of the CTD transformation, the atmospheric and illumination parameters varied in the physics-based model and demonstrates the influence the target manifold samples have on the orientation of the coordinate axes in the transformed space. A comparison between detection performance in the CTD subspace and spectral radiance space is also given for two hyperspectral images.