Why is constrained unmixing usually a bad idea?

The equations used for spectral unmixing (also known as spectral mixture analysis) can include constraints that the pixel fractions calculated for all endmembers be between zero and one (i.e., between 0% and 100%), and that all the endmember fractions for a given pixel sum to one. These constraint are popular because they result in endmember fractions for each pixel that seem physically reasonable. We tend to think that a value of zero means that the endmember covers less than none of the pixel, and that a value of one means that the endmember covers more than all of the pixel. 

Versions of ENVI before 3.0 did not support constrained unmixing because the developers strongly believe that constrained unmixing is unwise. It forces your results to appear reasonable, even when the endmembers are not reasonable. For example, if you use a sine wave, or a completely random spectrum as an endmember, the resulting endmember fractions will still fall between 0 and 1. If you always use unconstrained unmixing, then the endmember fractions will give you information about how good your endmembers are. If the endmember fractions are usually between 0 and 1 when the unmixing is not constrained, then the endmembers are probably appropriate for the scene.

Due to popular demand ENVI 3.0 and later versions do allow constrained unmixing. Nevertheless, the ENVI developers recommend that customers avoid using constrained unmixing so that they will be better able to evaluate whether their chosen endmembers are appropriate for their scene.

Often match filtering is a better choice than unmixing. It does not require accurate information about all of the endmembers in the scene. It can give information about the abundance of one endmember without information about the other endmembers. It is considered to be more robust for this reason.

Reviewed 10/15/14 by PS, KK