Soft decision-made hyperspectral measures for target discrimination and classification

Many hyperspectral measures such as Spectral Angle Mapper (SAM), Euclidean Distance (ED), Spectral Information Divergence (SID) calculate values that can be used to measure the closeness between two hyperspectral signatures in terms of spectral similarity. When a hyperspectral measure is used for target discrimination or classification, a commonly used approach assigns a data sample to a spectral signature with the maximum spectral similarity or a class whose mean is most similar to its spectral signature. As a result, such a hyperspectral measure is called a hard decision hyperspectral measure and its performance is evaluated by its confusion matrix. This paper develops a new class of hyperspectral measures, called soft-decision hyperspectral measures which use the similarity value between a data sample and a target signature or a class as an indicator of the likelihood of the data sample assigned to this particular signature or class instead of signature or class map resulting from hard decisions. In order for a soft-decision to perform target discrimination or classification, the soft-decision hyperspectral measure-generated likelihood values are normalized to probabilities so that a threshold can be used to make hard decisions via a recently developed 3D ROC analysis. Experimental study demonstrates that these two approaches yield different results.