Effects of Endmember Dimensionality on Subpixel Detection Performance

The importance of estimating the correct number of endmembers for hyperspectral subpixel detection applications is shown. To demonstrate this, various state-of-the-art methods to estimate the endmember dimensionality are identified. Additionally, a new method for estimating the number of endmembers is presented based on the specific subpixel detection algorithm used, the desired target signature, and the hyperspectral image being analyzed. A comparison of this new metric to current state-of-the-art methods shows gains in a number of experiments on real-world hyperspectral image data. Additionally, the experiments show how sensitive subpixel detectors can be to the number of endmembers when all other variables are held constant.