Hyperspectral Data Processing: Algorithm Design and Analysis

27.5 Selection of Reference Signatures

As noted, the selection of the reference signature vector r had significant impact on the performance. To address this issue, two general guidelines of how to select the reference signature vector r based on our extensive experiments may be helpful.

1. Since the signature vector r is used as a reference signature vector between two signature vectors s₁ and s₂, it should have some degree of correlation associated with both signature vectors. Keeping this in mind, when spectral feature characterization is performed such as signature discrimination via a database/spectral library, the best reference signature vector to be selected is the average of all signature vectors in the database/spectral library so that any pair of the two signature vectors drawn from the database/spectral library to be characterized can have some correlation with the averaged signature vector. On the other hand, if a reference signature vector is selected from the database/spectral library, the performance will be determined by how close a signature vector is related to the other signature vector coming from the same database, which is to be compared against. As a consequence, the performance may yield complete different results. Such evidence was demonstrated in Section 27.4.

2. By contrast, if a mixed signature vector is to be classified/identified through a database/spectral library which comprises each mixing component, then the best candidate for the reference ...