On the validation of a spectral/spatial CBIR system for hyperspectral images

We report validation results on a novel Content-Based Image Retrieval system for hyperspectral images that uses both the spectral and spatial features obtained from unsupervised unmixing processes. Spectral features consist of the set of end-members obtained from the image by an Endmember Induction Algorithm. Spatial features consist of statistics of abundance images. Both pieces of information are combined into a similarity measure defined between the hyperspectral images that guides the search for answers to the query. Such a system allows one to retrieve images containing materials similar to the query image, and in a similar proportion. We provide validation results using synthetic hyperspectral datasets.