HYPERSPECTRAL SPACE: TOWARDS A NEW INTEGRATED APPROACH FOR OBJECT-SPECIFIC CHARACTERIZATION

摘要

Remote sensing imagery is prone to illumination effects which reduce the performance of traditional spectral classification techniques. Using hyperspectral imagery we propose the new concept of hyperspectral space which allows one to visualize and quantify the recorded electromagnetic (EM) radiation in hyperspectral imagery as a new regionalized variable. Our goal is to present the underlying theory of hyperspectral space and to introduce a methodology which spatially describes the spectral properties of scene objects using the concept of hyperspectral space. This approach generalizes classical spectral theory by integrating explicitly the notions of spectral magnitude, wavelength topology and spectral scale as the basic components of object-specific spectral properties. The practical application of this theory is that current spatial tools can be applied to hyperspectral space to better recognize relevant landscape objects. In the context of this study, we use high resolution CASI imagery of the Duchesnay natural reserve in Ste-Catherine-de-la-Jacques-Cartier, in Quebec. Hyperspectral profiles are extracted from pixels of different scene objects and the variability of the spectral profile is characterized using the semivariogram. We conclude that the hyperspectral space concept may have the potential to facilitate the development of more adaptive segmentation algorithms which are less susceptible to sensing conditions effects in heterogeneous landscapes. The hyperspectral space concept brings a new dimension of information to natural resources management by providing new techniques to qualify and quantify parameters to be extracted from hyperspectral imagery.