Developing Spectral Libraries Using Multiple Target Multiple Instance Adaptive Cosine/Coherence Estimator

Traditional methods of developing spectral libraries for unmixing hyperspectral images tend to require domain knowledge of the study area and the material's spectra. In this paper, we propose using the Multiple Target Multiple Instance Adaptive Cosine/Coherence Estimator (Multi-Target MI-ACE) algorithm to develop spectral libraries that will capture the same spectral variability as traditional methods but require less processing time and domain knowledge. We compared traditional and Multi-Target MI-ACE generated spectral libraries' ability to accurately predict sub-pixel composition using Multiple Endmember Spectral Mixture Analysis (MESMA). Multi-Target MI-ACE spectral libraries maintained the same sub-pixel composition accuracy compared to traditional libraries, while significantly reducing model complexity. Additionally, the Multi-Target MI-ACE confidence values could be used to constrain MESMA model complexity and considerably reduce the number of endmember permutations needed. In summary, Multi-Target MI-ACE has been found to successfully develop spectral libraries that capture the full spectral variability compared to traditional approaches, while reducing MESMA model complexity and the need for domain knowledge.