Hyperspectral Unmixing and Band Weighting For Multiple Endmember Sets
Imaging spectrometers measure the response from materials across the electro- magnetic spectrum. Often, in remote sensing applications, the imaging spectrometers have low spectral resolution resulting in most measurements being mixed spectra from a scene. In these cases, pixels are assumed to be mixtures of pure spectra known as endmembers. Given the prevalence of mixed spectra, a common analysis technique is spectral unmixing. When the responses from materials are measured across hundreds or thousands of narrow contiguous wavelengths, then it gives rise to the concept of Hyperspectral Imaging (HSI).
The high dimensionality of the hyperspectral data results in highly correlated information in neighboring spectral bands. Also, irrelevant bands may degrade the performance of spectral unmixing methods. As a result, band weighting can aid hy- perspectral unmixing and endmember estimation methods. ‘Band Weighting’ refers to a method for assigning weights to each band in hyperspectral data that indicate the relative degree of importance during unmixing.
Many hyperspectral scenes possess multiple sets of endmembers. For example, consider a scene consisting of two distinct ground regions (eg. urban and rural areas) composed of distinct materials or endmembers. The image spectra from each region will then only be mixtures of the endmembers associated with the corresponding region.
The proposed methods perform endmember estimation for multiple sets of end- members, estimates proportion values, and assigns partition specific band weights. Band weights are estimated using weight decay and sparsity promoting techniques. The proposed approach is compared with state of the art method present in the literature. Results are shown on simulated and real hyperspectral data sets.
Khopkar P. (2014). Hyperspectral unmixing and band weighting for multiple endmember sets. Columbia, Mo.: University of Missouri.